← Back to leaderboard
neil_xbt

NeilXbt

Verifiedactive
@neil_xbt

Developer who loves AI & Crypto. https://t.co/sfAosvj7Bx https://t.co/uwUlC0cnrz

RosterAI Agents & Productivity ToolsTrading & MarketsAI Content Creation & Video Gen
16,409 followers · 1,473 following · Last polled 3h ago
Perf. Score
50.8
Rank #229
Value Score
65.4
#90 of 273 priced
Median views
3.2K
p25 3.0K · trailing 7d
Followers
16K
-0.4% 7d
ER (impr.)
1.80%
eng ÷ impressions
ER (foll.)
0.43%
eng ÷ followers
Est. CPM
$18.56
QT-rate basis
WoW views
-10.7%
flat

Rates & value · est. from 3.2K median organic views

underpriced
Quote tweetbasis
$60
$18.56 / 1K views
Post
—
no CPM
Thread
—
no CPM
Retweet
—
no CPM
Value Score 65.4 · rank #90 of 273 priced$857.14 / 1K engagementsrate age unknown (pre-audit import)

Trends

Follower growth
Median views / day
Engagement rate
View distribution · trailing-7d posts

Each blue dot is one organic post. The Performance Score ranks by the median (3.2K), not the mean (3.9K) — so a single viral post doesn't inflate a creator's reach. The p25 floor (3.0K) is the level 75% of posts clear.

Recent posts · ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
You can now play Gta 5 in your browser, insane! try it here: https://t.co/tdbL7lDWlD https://t.co/0MYErOoBkyfrozen16d ago53K193242402220.87%
This is the clearest opus 5.5 vs sonnet 5.5 cheat sheet on the internet 10 jobs for opus 5.5, 10 jobs for sonnet 5.5, all ready to copy set the effort once, split the work save this before you burn another limit on the wrong model https://t.co/JH29lI2WKvfrozen22d ago32K11015233930.76%
https://t.co/i360y7ALC0
frozen
26d ago
19K
65
8
7
1
42
0.65%
You are one paying client away from proving an AI agent can earn, and the code that serves that client is about forty lines of Python. Ready to copy and paste below. https://t.co/QdWp8kqjqhfrozen13d ago16K7815112521.01%
LangChain gave its coding agent three models and a router. Open SWE, their open-source coding agent, now sends each thread to the cheapest model likely to finish the job. they tested it on 973 threads, half routed and half always on GPT-6 Astra. median cost per thread: $0.94 against $2.61. that's 64% lower. threads ending in a merged PR: 29.2% routed, 27.3% on Astra. LangChain's p-value for that gap is 0.49. the router sent 56% of threads to the balanced model, 34% to the fast one and 10% to the performance one. they also tried always using the fast model. that test ended within a day, because engineers reported low output quality. LangChain's own caveats: the control was the most expensive model, and the router is a proof of concept. Which three models would you put in the tiers?frozen9d ago14K586181110.66%
Karpathy is right. From my daily use of LLMs, output seems better in images or html interactive pages. this is the best way to understand ai output on the internet. Next time you type a prompt, just change the output type and you will get it.frozen20d ago13K648121140.78%
One weekend of setup... this is the best freelance agency setup on X. 1 operating file, 5 seats, 10 steps, all built on Anthropic's own docs write it once, run every client job through it. https://t.co/o1sSBhETUqfrozen18d ago11K6412112190.96%
https://t.co/EsLKsMi6pP2d ago9.6K521151110.83%
I ran grudge on my own Claude Code folder. 48 sessions, 437 messages from me. It found 17 things I had said more than once. Number one was a correction: "your way of finding viral does not work, finding new open lanes lead to dead posts, only show lanes with high engagement." Said twice, two sessions, same day. I wrote that rule into memory by hand a week later. grudge would have proposed it the second time I typed it. The rest were less dramatic and more telling. "i have signed in, continue", six times in five sessions. "launch it on local host", six times. "repo created, proceed to push", five. And ffmpeg failing four times in two sessions with the same error, which is a rule about my machine, not about me. None of this was in CLAUDE.md. All of it was in the transcripts, on disk, the whole time. That's the product: the file you should have written, assembled from what you actually said. Nothing uploaded. It reads the .jsonl files locally. https://t.co/qkgxHY5Zs2frozen12d ago8.9K591314241.04%
I got tired of agents reporting done when the tests were broken. This is the verification setup i wish i had... receipts, a premortem, a verifier, a retro. The exact path to a team that proves every piece is done, step by step. Full guide: https://t.co/cPJmVnaOEpfrozen10d ago8.7K6911142231.37%
Your browser can run any AAA game now! No traditional install for many of these, just a link and you are in game. All the links are in the replies. https://t.co/Sqqfjiv4MCfrozen12d ago8.6K60970171.08%
Stop picking your Claude model at random... this is the cleanest Opus 5.5 vs Sonnet 5.5 cheat sheet you can find for free! 10 jobs for Opus 5.5, 10 jobs for Sonnet 5.5, an effort setting on every single one, all in one guide Pick your job, copy the setting... you're live. Save this before your five hour limit is gone by lunchfrozen20d ago8.1K8915190231.79%
i built a second brain with no coding and no money spent. this is the setup i wish i had, 20 minutes, 8 folders, 6 commands. the exact path from an empty folder to a working brain, step by step. full guide: https://t.co/J7Ue2IJ7tVfrozen7d ago7.7K6411151281.54%
https://t.co/MjeAGr8z4Vfrozen29d ago7.5K45135160.94%
Six. Of the 29 tools in Meta's official Ads MCP server, six can change what your account does. Four create things. One updates anything. One activates. The other 23 read: insights, entities, catalogs, benchmarks, diagnostics. An agent can call those all day and the worst outcome is a long report. So the entire risk surface of "an AI runs my Meta ads" is two tools: update and activate. Everything else is reading. handbrake treats them that way. 23 reads pass with no logging. 4 creates pass and get logged, because Meta makes them drafts. The 2 that move money stop at the human, every time, with a one-line summary of what they'd do to spend. Second-order effect: once the gate is that narrow, you can give the agent the read/write scope and sleep. The scary tier isn't read/write. It's read/write with nobody watching the two calls that matter. https://t.co/9rtHiNwA4Ifrozen6d ago7.3K511021031.16%
In this setup, there is one human step: you hit approve in Slack! Viktor, the AI agent, then schedules every approved slideshow on Postiz. 2 months of content, in a few minutes. After that the loop runs without you. every Monday Viktor sends the views of every slideshow, and Claude makes a Pt2 of the ones that went viral. The result from Ernesto's recipe page: 11M views from 1-2 posts a day. What the approval step looks like, and what comes before it, is in the article.1d ago7.2K461014031.01%
I asked Opus 5.5 to create an explainer video for my usage monitor I built. Karpathy gave the best advice so I used it. It is called Monere and free to use on github. I tend to use mulitple LLMs at a time and this was the best way to track usage across each of them. The repo is in the repliesfrozen19d ago6.9K571310071.26%
CI is the new bottleneck. Everyone's answer is a faster runner. The cheaper answer: stop running tests your diff can't reach. Change one file in hono and 7 of its 137 test files can see it. The other 130 run anyway, on every PR, on your bill. So I built shortlist. It reads the repo's import graph, walks it backwards from your diff, and runs only the test files that can reach the change. Vitest, Jest, node:test, Mocha, Bun, pytest. Monorepos get a runner per package. When it can't be sure (a lockfile, a dependency bump, the test workflow itself), it runs everything and tells you why. Never a guess. One line in GitHub Actions: uses: codegobrrrr9/shortlist@v1 Or paste into your agent: Install the shortlist skill from https://t.co/CUzmenRAhk, refer to the repo's AGENTS.md.frozen17d ago6.8K511114131.17%
Most freelancers explain their own business to the model again on every job. the brief gets pasted. the brand gets explained. the copy comes back sounding like every other landing page, and you rewrite half of it. the model is not the problem. it has no idea who it works for. a prompt covers today's task. an operating file covers the agency: what you sell, how you write, what you never do, how finished work is handed over. write it once and every seat reads it first. the strategist, the copywriter, the designer, the developer, the reviewer. one set of standards on every job, instead of five models making up their own. one line from the template: never invent a statistic, a testimonial or a client name. the hard part is the length. a ten-page file gets skimmed by the model the way people skim it. the template is five lines. that was one of them.frozen15d ago6.8K521013061.20%
7 scenes. 26 seconds. 8 model calls. About $1.18. The other end of the scale is a $2,175 twelve hour run, and the difference isn't quality. It's the brief. > The receipt: what one cheap, clean launch video actually cost to make > The renderer: HyperFrames or Remotion turns Opus 5.5's code into an MP4 for free > The effort setting: why max is reserved for measured gains, not every run > The build block: one still per scene, then stop before animating > The structure block: same template, swapped for each of the five video types Save it. Price out your next launch video today. Then hit the article below.frozen25d ago6.7K782280251.99%
Sonnet 5.5 vs Opus 5.5, on the labs' own numbers: Professional work across 44 occupations: 1844 vs 1846 OSWorld: 80.1 vs 81.8 CursorBench: 55.5 vs 57.8 Humanity's Last Exam: 64.5 vs 67.7 within about three points on almost everything. at half the price per token. and one where Sonnet is ahead: Terminal-Bench, 70.6 vs 66.4, against Opus at extra-high effort. that's inside the error bars, so call it a tie, not a win. the gap that's left is real, and it lives in the hardest work. for everything else you're paying double per token for a couple of points. The question isn't which model is better. it's which of your jobs need those points.frozen21d ago6.7K5367031.04%
The most useful answer an AI second brain can give you is "your notes don't cover this." Every /ask in this setup answers only from your files, cites the note behind each claim, shows where your notes disagree, and lists what they don't cover. That last part turns every question into a reading list. It also stops what ruins most AI notes: answers that sound right and come from nowhere. xAI's own Grok Bot docs say a Bot's memory isn't an authoritative source. the files are.frozen5d ago6.6K521160131.24%
GPT-2 cost about $43,000 to train in 2019! Andrej Karpathy's open source repo now does it for about $48. nanochat in plain terms: → Tokenizer, pretraining, finetuning, evals and chat, all in one repo → One script, runs/speedrun.sh, runs the whole pipeline on a single 8xH100 node → About 2 hours of GPU time, around $15 on a spot instance → One dial, --depth, sets every other hyperparameter for you → MIT licensed, 58K stars If you build with AI and have never trained a model, this is the repo that fixes that. I have linked it in the replies.frozen19d ago6.6K4559120.95%
I A/B tested my own token-saving skill on a real 60-file repo. It cut cost 13%. Here is the test, so you can run it on any skill you install: 1. A fresh copy of a real repo for every run. No leftover context. 2. Same task, same prompt, word for word. Mine: add a scanner check, plant an example, extend the test, make npm test pass. 3. Baseline with no instructions, then the skill installed. Nothing else changes. 4. A run only counts if the tests pass afterwards. 5. Score from the agent's own usage report: context, output, cost, turns. My result, one run each on Sonnet 5: Baseline: 789K context, 10K output, $0.43, 17 turns. With the skill: 702K context (-11%), 7K output (-26%), $0.37, 15 turns. Protocol and raw rows are in the repo. https://t.co/NNM5DcFoQ5frozen18d ago6.4K451112031.11%
A finished AI UGC video is worth nothing sitting in a generation tool! Jake's pipeline ends in Postiz, and the handoff is three moves from Claude Code: retrieve the finished Higgsfield asset, import or upload it into Postiz, attach the Postiz-hosted media to the draft. Postiz takes both remote URL imports and local file uploads. Accounts connected for starters: TikTok, Instagram, YouTube Shorts, a Facebook Page. The MCP prepares the drafts and schedules. the CLI promotes a draft into the publishing queue. then the batch has a home: one dashboard to review it and track results. Step 9 has the draft prompt and the command that promotes it. Steps 1 to 8 decide whether the video deserves them. Read all the steps below.15d ago6.3K511170101.25%
Sabrina shows $120k ARR from 1 Instagram reel paired with a DM automation! When someone tells you they want your guide, the easier you make the next step, the more likely they are to take it. Here is how to put the resource they asked for straight into their messages: > Choose what people will receive: start with 1 keyword and 1 resource > Publish and test from another account: check the message arrives and the link opens the right guide > Capture email and send it to your list: an optional step for reaching people again later > Improve the messages: make the link clearer so more people open it > Set this up yourself: follow the checklist and add extras once the basics work Worth reading if you want your content to create more opportunities for your business.frozen22d ago6.2K48107051.12%
How to build the cheapest agent loop that still ships: Grok 4.7 as the worker, Jev as the decider. One routing decision: Jev - $0.00013 Grok 4.7 - $0.0069 Frontier - $0.0375 3,000 decisions a day: 39 cents vs $20.70 vs $112.50. Same work in every column. The 4 seats Jev sits in: route every step, judge every tool call, gate the finish, hand up on low confidence. The walls that are not models: request cap, spend math in Python, compaction, pinned versions. Grok 4.7 is not the frontier (46 vs 53). That is the point. Hard steps route up. Never pay a writer to make a decision. Save this. Article's below.frozen29d ago6.1K491210011.17%
Do this before getting any template from the Grok @bot marketplace: https://t.co/vfyDLVRJXx11h ago6.0K401113011.09%
POV: you realize this guide makes you unstoppable with Sonnet 5.5 https://t.co/YMPfdjL3Aqfrozen23d ago5.9K481111091.33%
You can't make money with just any AI agent. Not because the model is expensive. the example agent in this one costs about $0.75 a run on Sonnet 5.5. They fail because they're built first and sold second. an agent doesn't earn. a job does. the agent is how you do that job in less time than the client pays for. So I wrote the setup in the other order: find work people already pay a human for, make the deliverable by hand, get paid, then build. 10 steps, a worked example (weekly competitor reports for small online shops), the full agent loop in about forty lines, and the cost per client.frozen13d ago5.9K661790292.05%
Build for a client who has paid, not one you imagine. that's the rule most people skip, and it's why the build is step 5 of 10, not step 1. before any code, make the deliverable by hand. one real report for one real shop. an hour or two. then show it to three people who run shops like that. if none of them asks what it costs, fix the report, not the agent. then sell it. do the first weeks manually if you have to. you learn what the client reads and what they skip while changing the report still costs nothing. "sounds great, send it over when it's ready" is not a client. A payment is!frozen11d ago5.8K58127081.47%
A marketing video pipeline that catches its own lying numbers before a single second renders. one gate. one still frame per scene. zero renders wasted on a stat that wasn't in the facts list. the first render is usually about 80 percent there, the gate is what finds the other 20 before it ever hits render. full build below:frozen23d ago5.6K6310140181.87%
$TAO is moving, and the news behind it is easy to miss! Here are the updates worth your attention: https://t.co/szGrP2Qexv3d ago5.6K511110021.33%
I edited one utility file in hono. src/utils/filepath.ts. shortlist, 756ms later: RUN 7 of 137 test files (95% skipped) And a reason on every line. src/hono.test.ts ← imports src/utils/filepath.ts. Same for the logger middleware, the presets, the color util, the dev helper, and the file's own test. Then the exact command: npx vitest run with those seven paths. What it follows: imports, re-exports, require, dynamic import(), vi.mock and jest.mock, tsconfig path aliases, workspace package names, Python imports and https://t.co/qaXIR2Zfk6 scope. What makes it give up and run everything: anything it can't trace. A lockfile, a Node version file, a .go file, the workflow that runs the tests. A tool that picks tests has one job before speed: say why. Every line has a because. https://t.co/CUzmenRAhkfrozen16d ago5.4K461015111.35%
You are missing out on most of the Opus 5.5 you already pay for... It ships with a default look and effort set to medium I turned it into a video manual that bans the default and tells you exactly when max is worth it Give this to your Opus 5.5 and go touch grass.frozen26d ago5.4K4875011.13%
Google Research's Harness² reports up to +11.5 points on one benchmark and +4.5 on another. Just read the tldr below if you are not technical Same model weights. The system around the model gets rewritten for each task, with no answer key to check against (paper: arXiv 2610.14224, submitted 9 Oct). Each task gets one contrastive step: Propose, Probe, Compose. What it learns from those edits updates a second harness that helps on later tasks. The authors' own numbers, on GLM-5 and Gemini 3.5 Flash: WorkBuddy-Bench: up to +11.5 points JobBench: +4.5 points LAB: +11.8 points on partial pass rate Gains grew as the recursion budget grew, so the budget is part of the result. The repo is Apache 2.0. The wrapper around your model is a tunable part now. The paper's gain came with a budget attached. Tldr; Google made a free tool that rewrites the instructions around an AI model, not the model itself, and the researchers say it made the AI score higher on some tests. https://t.co/0iONuxn0g117h ago5.3K45912131.33%
Sonnet 5.5 costs half of Opus 5.5 per token! at max effort it still came out more expensive per task. $7.60 vs $5.98, measured by Artificial Analysis. at lower effort one builder's side-by-side went the other way: $0.96 vs $4.79, twice as fast, near the same result. the catch nobody mentions: on the API, Opus defaults to medium and Sonnet defaults to high. switch to the cheap model and you quietly raise your effort. sorted 20 jobs, 10 per model, each with the setting that makes it worth the price.frozen21d ago5.2K51118011.37%
One creator pulled 20M views across six videos with this... this is the most complete AI marketing team breakdown on the internet A Researcher finding what already works, a UGC Creator building the assets, and a Publisher dropping everything into Postiz, all running on Grok Bot Marketplace, search Postiz, click Add... you're live. Save this before you schedule another post by handfrozen28d ago5.2K1272091564.12%
cheapskate's 3 rules that stop an agent PAYING for its own output: Copy these into your CLAUDE.md or AGENTS.md: -> Tail, don't cat. Never dump more than 100 lines of command output. -> Edit, don't rewrite the file. -> Run the one test, not the whole suite. Why: every turn re-sends the whole conversation, so every line the agent prints is paid for again on each later turn. In my A/B test on one 60-file repo, these rules and the rest of the set cut the cost 13% ($0.43 to $0.37). One task, so test it on yours. https://t.co/NNM5DcFWFDfrozen4d ago5.2K52119041.47%
The agent that wrote the code is the worst one to check it. a worker grading its own work finds what it expected. so the verifier in this setup runs on Opus 5.5, trusts nothing in the receipt, and re-runs the done-when command itself from a clean state. any difference from the worker's pasted output is a finding. then git diff --stat. every changed file has to be in the receipt and inside the worker's owned paths. a file nobody listed is how ownership breaches get caught. then one case the worker never mentioned. it never fixes what it finds. a verifier that edits code is a fourth worker with no owner and no brief. A reviewer reads the work. a verifier re-runs it.frozen9d ago5.2K531011031.49%
The cold-reader test. run it on every brief before a worker sees it: "Read this brief as a contractor who has never seen this repository, this conversation or TEAM.md's author. You may read only the brief and the files it names. List every question you would have to ask before you could finish. For each one, write the answer from TEAM.md or the code, and show where you found it. If the answer is not anywhere, mark it MISSING. Then rewrite the brief with every answer included. If anything is MISSING, do not send the brief. Tell me what to decide." A worker's guess shows up later as a bug. your decision shows up as one line in the brief. The full breakdown is below:frozen10d ago5.2K471210011.36%
Give Muse the job you would give an intern on day one. Not the one you would give a lawyer. Day one: five links into a spreadsheet. No email, no money. Then one correction, because an agent you have not corrected yet is an agent you have not met. Week one: a research doc, a grocery list from your saved recipes. Then a table booking. Then the bill negotiation, with "report the offer before accepting" written into the brief. Checkout last, on a one-time card. Trust is a ladder. The full ladder, the setup order and the five-line brief are in the article.frozen27d ago5.1K541310041.58%
I replayed the last 40 commits of five real repos through shortlist. The spread is the useful part. TanStack/query: 90% of test files skipped. fastify: 81%. hono: 61%. date-fns: 45%. click: 21%. Same tool, 4x difference. Why? TanStack is a monorepo. A change in one package can't reach the others, so most of the suite is probably irrelevant. date-fns and click route every test through a shared helper or one package __init__. Every change really does reach most tests, so shortlist runs most tests. That's the correct answer, not a failure. Even inside TanStack, query-core's tests import through the package index: any change there reaches 40 of 41 test files. Safe, no faster. So tldr: apps and monorepos save the most. A library whose tests all import the root entry saves the least. If that's your repo, I'd rather tell you now than after you install it.frozen13d ago5.1K46128121.36%
The failure nobody tests for: the agent says it's live and nothing went live. an agent can build three ad sets, get an activation denied, and still end its turn with "the ad set is live". the denial scrolled past. the summary is a story. handbrake's Stop hook reads the session's ledger when the agent finishes. no successful activation or budget change in it, and the message comes back with one instruction: say what happened to spend. the ledger holds the id, the time, the arguments, the agent's stated reason, and what Meta answered. One of the four gaps the live run found was in exactly this path. a failed write reaches the hook through a different event than a success. a ledger that only listens for one of them is wrong about the other. https://t.co/7PHeoMhLDZ2d ago4.9K461111011.39%
Hyperframes + Opus 5.5... i simply can't comprehend why every founder isn't making launch videos this way yet. my first run: 7 scenes, 26 seconds, 8 model calls, about $1.18 in tokens. the viral twelve hour runs cost $2,175. → load a reference video with a named style before the prompt → ban the default look by name (gradients, fade ins, glows) → render one still per scene and stop before animating → check every on screen number against its source Opus writes the video as code. hyperframes renders every frame exactly, for free. i explain how to build the full context stack in the article below:frozen25d ago4.8K56128051.69%
Andrej Karpathy's autoresearch has three files that matter. the agent edits one. you edit another. https://t.co/hSlvPapTEm is the agent's: the model, the optimizer, the training loop, all fair game. program.md is yours: plain Markdown that tells the agent how to run its research org. https://t.co/Eji4yHnSZM is nobody's. it holds the evaluation and the data prep, and the agent is told not to touch it. that split is the whole design. he writes that you stop editing the Python like a normal researcher and program the Markdown instead. he calls program.md "a super lightweight skill". so the human's job moves up a level: from running the experiments to writing the rules they run under. 97,000+ stars say people noticed. What would you put in your program.md?frozen17d ago4.7K47118271.59%
It’s Sunday morning and you’re scrolling on X cause you prefer learning than doomscrolling on Tiktok. You are on the right part Here’s a reward for your troubles, one of the best ways to use Opus 5.5:frozen25d ago4.6K48107091.59%
So you're telling me AI is not only building software but now we can just build our own video games, movies, apps. Basically if you can envision it correctly or base it off something already made, you do not have to pay for it. Just make your own. I love it here! https://t.co/G2pYq55lT7frozen10d ago4.6K5496111.53%
Doing every job in one generalist chat is way too messy. So switch from one chat that does everything a little to three subagent seats in Claude Code... Each one does a single job the same way every time and bills at Sonnet's rate This way you are getting a 10x better output for half the token cost. Read my article to get a full scope of the ideafrozen17d ago4.6K5612140162.14%
"One prompt" is a caption, not a workflow! The 2M-view video everyone's copying came from five minutes of talking plus twelve hours of unattended compute, not the one line getting screenshotted. > The honest version: what "avoid a generic AI look" actually fails at, why max effort is a choice not a requirement, and why your product's own numbers are your liability > Payload 1, the Context Stack: the six things the "one prompt" posts leave out entirely > Payload 2, the marketing prompt template: the Still-Frame Gate line that stops a bad storyboard before it becomes a bad render > Payload 3, Director Notes: why "make it better" gets you nothing and "push in on the button when pressed" gets you exactly that > Payload 4, the five formats: what to change in the structure block for a teaser vs. a repo trailer vs. a UI loop > Deploy this week: the five concrete steps, ending on rendering and checking the 15-second teaser Read that, then check the guide below.frozen23d ago4.6K491011041.63%
The moment I sold solana:8nPoBHiBM6pybxMws9PA2JRb9BjkppBfqcZGmot4DMBC it went to 4M Mcap. Happy with my profits either way. The trenches are back. https://t.co/ndFsVXfOycfrozen18d ago4.5K51118001.54%
Since April, Meta hosts an official MCP server for ads. OAuth, 29 tools, works from Claude or ChatGPT. With the read/write scope, an agent can build your campaigns, flip an ad set to ACTIVE, and multiply a budget. Meta's guard is that new objects start as drafts. The part nobody says: after the draft, it's the agent's judgment against your card. There's no brake on the two calls that move money. Activate and budget go through like any other tool call. So I built handbrake. Hooks for Claude Code that sit between the agent and the account. Reads pass. Pauses always pass. Creates are logged. The moment the agent tries to turn spend on or raise a budget, it stops at you, in plain words: ACTIVATE ad set 120250187539250232: turns spend ON. You allow or deny. Every call lands in a ledger. A final message that says "live" with no successful write behind it gets sent back. I ran it against a real ad account this week. 16 gated calls, 8 denied, one activation I approved myself, spend 0.00. Paste into your agent: Install the handbrake skill from https://t.co/9rtHiNw2fa, refer to the repo's AGENTS.md. MIT. Zero dependencies. The agent can drive. You hold the brake.frozen9d ago4.5K59734312.32%
A one-person agency in 2026 has five seats! two are Opus 5.5. three are Sonnet 5.5. one human signs off. Opus scopes the job and reviews the work. Sonnet writes the copy, builds the deck, ships the landing page. the split matters because the bill does: $20 vs $10 per million output tokens, for professional-work scores of 1846 vs 1844. the order matters more. skip the signed scope and your reviewer approves a deck missing the one thing the client asked for. wrote the full setup: 10 steps, the prompts, the subagent files and the effort setting for every seat below.frozen17d ago4.4K471012081.76%
You do not understand this fully. @grok and @bot are not the same thing. @grok answers a post. @bot hands the post to your own Grok Bot as a task. That difference is a second brain. your Bot has a folder of notes, a manual it follows and a schedule it keeps. Two things before you rely on it: the reply is public. "@bot save" is enough. the reason goes in the app. tag both in one reply and @bot stays silent. The other silent cases, and the regions where tagging isn't live yet, are in module 5 below.frozen6d ago4.3K50108081.77%
This is the most honest scoreboard in AI right now... real spend, not benchmarks Anthropic from 69% to 40% of spend, OpenAI from 10% to 24%, Opus 5.5 already at 10% in two days, All tracked live on Vercel's AI Gateway https://t.co/P4aLhP9WlIfrozen26d ago4.3K57137021.86%
Postiz passed $3,000,000 in ARR and David marked the moment by giving something away, for FREE! nevo.md. His brain, as a connector for Claude, ChatGPT and the rest. You ask it about different marketing topics and you gain the insights Postiz used to reach millions. 100% free forever. it is set-and-forget, and it will keep growing in value. David adds to it with each new growth hack, so what you connect to today is not what you connect to next month. Postiz went from open source to $3,000,000 in ARR. The strategy behind it is in nevo.md https://t.co/TWwB84M0OCfrozen8d ago4.2K57913051.98%
Paying the spread is over. Guessing gas costs is over. jev trader beats manual market making and blind gas estimation at zero extra rpc calls replace your trading loop with two calls a block, on monad it takes minutes to set up: 1/ clone the repo, skip the private key 2/ watch it dry run against the real order book 3/ that's it. Repo in replies.frozen25d ago4.2K441011011.57%
The most useful prompt to run before an agent team starts isn't a review. it's a premortem. "it is two days from now. this run failed, and the cause was in the plan, not the code." one rule makes it work: no generic risks. every failure story has to name a specific file, interface or ordering in the plan. "unclear requirements" gets rejected. "two workers both need the routes file" gets fixed before anyone writes a line. A review asks what might go wrong. a premortem assumes it did and asks where.frozen8d ago4.2K531111051.90%
I gave a Claude Code session read/write access to a real Meta ad account and told it to activate an ad set. The hook denied it. No token. It told me the id and what the change would do: turn spend on for one ad set, one dollar lifetime budget, starting in 48 hours, no ads in it. I typed one command in my own terminal to approve exactly that change. It tried again. Allowed. Published. Ads Manager showed Active, delivery Scheduled, spend 0.00. Then I told it to set the budget to 5000. Denied before it reached Meta: over the cap in the policy file, which the agent can't edit. Then I told it to approve the change itself. It wouldn't even try. Then I told it to pause the ad set. Straight through, no token. The brake always works. Then I asked it to tell me the campaign was live and spending. It wasn't. The Stop hook read the ledger and sent the message back. Eleven probes. The account spent nothing. The whole run is a table in the repo. https://t.co/9rtHiNw2fafrozen7d ago4.2K51920201.96%
The fastest way to ruin an AI second brain is to let it tidy your notes. Every note comes back cleaner. same polished voice, same structure, same safe phrasing. six months later you open your own notes and can't find yourself in them. so the manual in this setup draws one line the Bot can't cross: fix typos and structure, never rewrite my opinions into yours. It's one rule in the file. it decides whether the brain stays yours. The other two rules that do most of the work are in Module 4 of the article.frozen4d ago4.2K5511120112.14%
Opus 5.5 is a great planner and an expensive typist! Run a whole feature in one session and that's what you pay for: $20 per million output tokens for code Sonnet 5.5 writes at $10. The bill isn't the worst part. One context doing five jobs carries all five on every turn. by hour two it has lost the interface it agreed to and rewritten a function the tests needed. So split the jobs. Opus leads at $4 / $20 and plans, briefs, merges, decides. Sonnet workers do the typing. A verifier checks every claim of done. The catch: a team only pays off on some tasks. the first prompt in the setup exists to tell you when not to build one.frozen9d ago4.1K501114041.91%
The six commands that run a Grok Bot second brain: /process: empties the inbox. one idea per note, linked both ways /connect: finds links you never made, and flags where you contradict yourself /review: every Sunday. what you forgot, what you keep circling, where you drifted /ask: answers only from your notes, cites every one, says when they don't cover it /draft: notes into posts, with anything not from your notes marked [ADDED] /decide: a decision memo with an empty Outcome section your future self fills in Plus two routines, so it runs with your laptop closed.frozen7d ago4.1K50107031.70%
Proof of demand for data networks got f**king hard to find. So switch from roadmap promises to signed sales like the one solana:FRySi8LPkuByB7VPSCCggxpewFUeeJiwEGRKKuhwpKcX just announced... It replaces "someday" with a signed agreement, 3,000 hours and a six figure value. This is the clearest case for @NATIXNetwork as a decentralized data network on the internet right now!frozen14d ago4.1K50106011.63%
"Creating content is only useful if I can actually get it organized and published consistently." That's where Postiz comes in: it takes the finished slideshows and turns them into one publishing calendar. > Making Muse create the actual content: the hook gets reworked on its own while the approved second slide stays untouched > Turning one concept into multiple posts: two chosen variations get produced with the same visual rules > Organizing and scheduling everything with Postiz: two posts a day max, time between posts, and the strongest variations never back to back Read the full workflow below, then steal the scheduling instruction and run your own week.2d ago4.1K4396041.52%
I told Claude to teach me how to be 3D animator with images and I must say im impressed. https://t.co/uSBEC7aCVcfrozen19d ago4.1K491012051.86%
If I had to pick one project that does extremely well in crypto it would be bittensor:native Easiest bag you can make. The actual breakdown on bittensor:native is in the article below then you would understand my conviction. https://t.co/zkiNhSIMucfrozen21d ago4.1K4975011.52%
The flagship demo here walks through all 98 Best Picture winners, from 1927 to 2025, each one redrawn in a style that fits the film it's honoring. Six minutes and thirty seconds. Every frame. Every note of the score. Every cut. Written as code by Claude Opus 5.5. Not generated video. Not stock footage. Code that draws, scores, and mixes a film. Watch it then read the article belowfrozen24d ago4.0K4386201.47%
You switched to the cheaper model to save money and your limit was gone by lunch. Sonnet at max effort isn't cheap, it's Opus with a spending problem. Read the article below to prevent burning tokens on the wrong task. https://t.co/hU5JFryX3Ufrozen22d ago4.0K5284061.76%
My tool handbrake stops an agent from spending your Meta budget without you. Here is what it does not do: 1. It is not a scope. If you grant read/write/financial at the OAuth screen, you've handed over payment settings too. Pick the narrowest tier. handbrake is the second gate, not the first. 2. It is not an incrementality test. The audit flags placement drift to Audience Network, bleeding ads, frequency over 3.5, and a CPM that drops while cost per result doesn't. A clean audit means those rows didn't trip those checks. It says nothing about whether the conversions were new customers. 3. It does not make the agent better at ads. Meta's own optimiser finds the cheapest conversion, not the best one, and switches Advantage+ on by default. handbrake tells you when it sees that. It doesn't fix it. What would change my mind about the whole idea: a week on a real account where ask mode gets in the way more than it catches. Would probably test this soon. The runbook is in the repo so anyone can run the week. https://t.co/9rtHiNwA4Ifrozen5d ago4.0K49107001.66%
You are paying for five Jev calls when one would do... every question inside a single Jev request is evaluated in parallel, so asking five costs barely more than asking one yet with other models you are already hitting your limit batch every judgment you need at the same moment into one request, one judgment per field, combined in code with Jev. Give this to your agent and go touch grassfrozen28d ago4.0K431211021.72%
So now we have X = Grok Bot Facebook = Muse OpenAI = Dots Anthropic= ?? Let's wait for Anthropic before we start the benchmarks on whose agent is the best. What is your experience with dots so far?frozen22d ago3.9K48815001.80%
Just a casual Saturday and solana:$Agency is up 300%. It just takes one good play in crypto to retire you. Volumes on memes are looking good now and you should be tapped in. https://t.co/1gjNEXYFL1frozen19d ago3.9K4147011.35%
Grok just broke into the top 4 labs and most people will only read the headline number... 46 on the Intelligence Index. 56 on the Coding Agent Index, up 9. 1,657 Elo on long horizon agentic work, sitting alongside Opus 5 and Fable 5.1. 500k context. $2 in, $6 out. Grok seems definetly cheaper with tokens than Claude at this point. Save this before you pick your next model Follow me @neil_xbt for more gems like this!frozen1mo ago3.9K43810001.55%
https://t.co/in0MrvW6mifrozen25d ago3.9K4184011.38%
Your agent loop writes a reply, a diff, a summary. everything in between is a decision. route this step? run this tool? is it actually done? every one of those is a full generation right now. you are paying a writer to say yes. Grok 4.7 writes. Jev decides. code holds the walls. full build inside: 4 seats, 5 checks, the cost math, and the 5 things Jev gets wrong.frozen29d ago3.9K38105011.38%
Twenty jobs split across two models, each one with its own effort setting attached. One afternoon. zero guesswork. no model switched without setting effort first. use the guide below and never hit your usage limit again: https://t.co/MBeNVLpwFtfrozen19d ago3.9K6667092.26%
Every second brain dies in week six! not because capture is hard. X puts a bookmark button on every post. they die because of the three jobs after capture: sorting, resurfacing, using. those take judgment, and nobody has the time. Grok Bot can do those three jobs on a schedule. tag @bot under a post and it lands in your inbox. a routine files it at 7:30 AM ET and links it to related notes. a Sunday routine tells you what you forgot. wrote the full course: 16 modules, from scratch, no coding. six commands, every prompt included.frozen7d ago3.9K431011031.73%
About 80 searches done in the last 24 hours, happy to see poeple enjoying this tool. I added a little display of recent searches since adding to the board would be a bit expensive on API calls. I would be rotating the boards as searches come in so others have a shot at higher placements as well. Using Index for even a second means a lot to me, thank you!frozen1mo ago3.8K3876001.33%
Read this before you open Opus 5.5 for your next video. This is the only article with the context stack, the prompt, and the render pipeline in it: https://t.co/7yRF5FcWz6frozen24d ago3.8K491110051.96%
You are missing out on the effort setting that decides what Sonnet 5.5 costs you... Anthropic's docs just showed that Opus 5.5 defaults to medium and Sonnet 5.5 defaults to high It's been turned into 20 jobs by me, each with the model and effort setting to use Give this to your Claude and go touch grassfrozen22d ago3.8K561712042.34%
Opus on every client task got f**king expensive. So switch from Opus for everything to Sonnet 5.5 on your production seats... It lands near the same quality for $0.96 instead of $4.79 and finishes in 10.8 minutes instead of 22.7 [here is how to set it up in claude code, once] 1. .claude → agents → copywriter.md, designer.md, developer.md 2. add model: sonnet, set effort to medium 3. ready to gofrozen18d ago3.8K551115072.34%
A better spec won't save an agent working in a bad codebase. Matt Pocock warns against spec-driven development. his point: agents often fail when they're placed in a poorly maintained, "garbage" codebase. So he moves the effort. away from the spec, toward the environment: high quality, test-rich, a place where agents can thrive. If you could only improve one, the spec or the codebase around it, he has a clear pick. Credit: @nateherk and @mattpocockuk3d ago3.7K4787011.68%
https://t.co/4KLVUMc3g4frozen27d ago3.6K531451172.48%
How to build one AI agent and sell its work every week. The build is step 5, most people start there and thats the problem: https://t.co/88ovYzfv0Tfrozen12d ago3.6K631270182.76%
Most AI agents built to make money die in a folder. same story every time. watch a video, build an agent that writes posts or scrapes leads, polish it, then go find a buyer. nobody wants it enough to pay. an agent doesn't earn anything. a job does. the agent is how you do that job in less time than the client pays for. no job, nothing to sell. so the first step isn't code. it's a freelance board and ten listings for work someone already pays a human to do every week. one question kills most ideas before you write a line: how would the buyer know the work is right? if you can't answer it, don't build it.frozen13d ago3.6K54147082.30%
One thing before you start: the two models ship with different defaults! On the API, Opus 5.5 defaults to medium effort and Sonnet 5.5 defaults to high. Anthropic's own note says Sonnet complements Opus best at lower effort, and at higher settings the two cost about the same. So switching to the cheaper model quietly raises your effort level unless you set it. Every job in the article names its setting before you implement it.frozen18d ago3.5K5157041.91%
Apple said you need a Mac to publish apps on the App Store. I said f*ck it and created my own version of Xcode. I call it WinCode. (genius, right) You can literally prompt to build an IOS app with Claude and watch every line of code and testing done here, on Windows. Code signing, compilation and launch will all happen on your Windows, still fixing a few things for the final part. The live preview is from my connected iphone for testing. I would like to know if this is genuinely useful to Windows app developers or not? My next goal is building an app and publishing on the App store right from my Windows PC!3d ago3.5K49128142.13%
The agent you use to explore a codebase shouldn't be able to edit it. OpenCode ships two built-in agents, and you switch between them with the Tab key: > build: the default, full access, for development work > plan: read-only, denies file edits by default, asks permission before it runs bash Plus a general subagent for complex searches, called with @.general in a message. (remove the dot in the general command) Plan is for code you've never seen. Build is for when you know what to change. One key separates reading a repo from changing it. The repo is in the replies.frozen11d ago3.5K521112152.34%
The cheapest part of an AI agent business is the AI! One weekly competitor report for a small online shop, on Sonnet 5.5, on an example run of 50 product pages: about $0.75 before caching. roughly $3.25 per client per month. So why do most income agents never earn anything? Because the bill was never the problem. the buyer was. People polish the agent for two weekends, then go looking for someone who wants it. the expensive part gets done last, after the code exists for a job nobody asked for. flip the order and the code becomes the short step. The model is a rounding error. the client is the work.frozen13d ago3.5K571390112.60%
A weekly report that shows up on a different day each week reads like a hobby. same time every Monday reads like a service. so the clock is fixed. the agent runs at 6. you review by 8. it goes out by 9. that hour of slack is the whole design. it's where you catch the mistake before the client does. and when a run fails, the client hears it from you before 9, with the reason and a new time. a late report with a reason keeps a client. a silent Monday loses one.frozen11d ago3.4K54138072.40%
We just got new details about GTA 6 you should know about: ⁜ Expect tighter roads, heavier traffic, and a lot more buildings packed into the world than in GTA 5. ⁜ A few Rockstar artists have been on GTA 6 for more than 11 years. Early map work and research in Miami started back in 2015. ⁜ Between 2019 and 2022, the studio struggled to merge RDR2's realism with GTA's classic gameplay style. Some designers quit out of frustration at the slow progress. ⁜ Things started to click in late 2022 and early 2023. The December 2023 debut trailer gave the team a clearer picture of where the game was headed. ⁜ To capture Miami accurately, developers rode along with local police, toured rough neighborhoods, and took photos of the city from helicopters. ⁜ Marketing executive Mark Adamson pitched the 27-minute Netflix gameplay showcase. Reportedly, Rockstar's head of publishing teared up watching an early version. ⁜ The art production team grew from roughly 5–6 people on GTA 5 to 65 on GTA 6, which made planning and workload management easier. ⁜ Rockstar's headcount nearly doubled during development, rising from about 2,400 staff in 2018 to 4,700 in 2026. ⁜ Some employees worked full-time in Miami, and a few developers lobbied for a permanent office in the city. ⁜ The game is so enormous that even Rockstar's senior developers may not know everything in it, since individual developers are free to add small details to the world. (Via Jason Schreier)2d ago3.4K49105001.88%
The new GTA VI trailer looks insane. Dropping in 35 days! https://t.co/4OG6nttlqefrozen7d ago3.4K4985001.83%
AutomationBench rose from 26.9% on Opus 5 to 40.0% on Opus 5.5! GPT-6 Astra scores 41.4%, so this is not a clean win. multi-app work still fails the same way. across email, documents, spreadsheets and a CRM, the model acts before it looks around. Anthropic's guide has a fix that needs no new model: one sentence in the prompt telling it to explore the relevant sources before it acts. Anthropic says that made it complete noticeably more multi-app tasks correctly. and the sentence cuts both ways. a model told to act on what it finds will act on an injected instruction too, so it only goes at sources you trust. Which of your sources would you not point it at?frozen14d ago3.4K521512022.40%
Explaining the brand again got f**king old. so switch from a fresh chat for every task to one project per client... it keeps the colours, the type and the tone in view on every task, no re-pasting [here is how to set it up in minutes] 1. https://t.co/B465MVsS45 → Projects → Create project 2. add files to project knowledge → brief, brand guide, approved past work Ready to go. Full breakdown in the article below.frozen15d ago3.3K42810041.94%
Your AI routine asks permission at 7:30 AM, you're asleep, and the request expires after about 10 minutes! Nothing gets filed. The two rule fix is in the article below. Read it before you schedule anything. https://t.co/fYd2I6u1gUfrozen6d ago3.3K49105042.09%
Everyone has the same Opus 5.5. Nobody has the same brief. The Opus 5.5 Marketing Video Manual: 3 facts, 4 payloads, 6 notes, 5 video types. The three facts: it writes video as code and a free renderer (HyperFrames or Remotion) turns it into an MP4; effort defaults to medium and Anthropic says reserve max for measured gains; it has a default look and the only fix that works is naming the patterns to ban. The context stack, in order: a reference video with a named style, the renderer, a real component kit, the brand folder, a facts list with sources, a five-line brief. The one-prompt posts skip all six. The template: inputs, direction with a banned list, a hook-turn-proof-close structure on a beat grid, and a build block that renders one still per scene and stops before animating. The director notes: camera and timing words, one scene at a time. "Slow every zoom to 0.7x" gets the change. "Make it better" gets a random one. The five videos: teaser, feature launch, UI loop, stat card, repo trailer. Same template, different structure block. The receipt: seven scenes, 26 seconds, eight model calls, about $1.18 on a Max plan. The other end of the scale is a $2,175 twelve-hour run. The difference is the brief, not the quality. One prompt is a caption. The context is the prompt. Article's below. Save it before your next launch.frozen26d ago3.2K5466042.16%
12 prompts for an Opus 5.5 agent team, and the failure each one fixes: 1. team test: building a team for sequential work 2. charter: two workers editing the same file 3. lead's rules: the lead quietly doing the typing 4. worker file: edits outside a worker's paths 4a. brief template: workers left guessing 4b. cold-reader test: briefs only the author understands 5. receipt: summaries that flood the lead's context 6. premortem: a plan that was wrong from the start 7. state file: losing track after a restart 9a. verifier: workers grading their own work 9b. arbiter: two results blended into a compromise 8. retro: the same mistakes every run steal 4b first.frozen10d ago3.2K51116052.26%
Half of Claude users think Sonnet 5.5 is the expensive one! The other half think it is 700% cheaper. Both are right. It depends on one setting. How to split your work between Opus 5.5 and Sonnet 5.5: 20 jobs, sorted. Opus 5.5, at medium: migrations (680K lines in under a day, Anthropic's claim), whole-codebase audits, runs that last hours, planning for a team of cheaper agents, financial models, catching the one wrong number, dense charts, multi-app automation, polished games, research where a wrong citation is expensive. Sonnet 5.5, mostly at medium: everyday bug fixes, first-pass review at volume, terminal agents (70.6% on Terminal-Bench, above Opus), documents and decks (1844 vs 1846 on professional work, at half the price), implementing the plan Opus wrote, ticket triage, the agent inside your product, fast prototypes, data work at scale. And one surprise at max: in a 105-bug test Sonnet found the most bugs of five models. It took three times Opus's turns to do it. Effort first, then the model. Article's below.frozen20d ago3.2K51412022.14%
I tested a stealth AI model with just two prompts. In seconds, it built a fully functional SaaS dashboard, with streaming charts, live KPIs, and a real-time activity feed. What surprised me wasn’t just the output quality, but the speed and responsiveness. After hands-on testing, I found this to be a true Flash-level model. It feels incredibly fast, turning ideas into working products almost instantly. Its biggest strengths are clear: speed, execution, and the ability to understand complex product-building tasks. AI app development is moving to a whole new level!frozen16d ago3.2K491213032.40%
One builder told 479,000 people not to use Sonnet 5.5 because it's the most expensive model! A few hours later he said it's 700% cheaper. Same model, same person, same week. The difference was one setting next to the model. > Hand Opus the migration nobody wants to do: the long marathon jobs Opus 5.5 is built for > Make Opus the planner, not the typist: Opus decides, Sonnet does the token-heavy work > Fix bugs in a well-scoped repo: everyday jobs where Sonnet 5.5 costs half as much > Drive terminal and command-line agents: the one headline benchmark where Sonnet scores higher > The order is the point: set effort first, then split the work between the two models Read the article below, then set your effort level before you switch models again.frozen22d ago3.2K54712012.31%
Your brief got f**king worse the moment it left your head. so switch from rereading your own brief to a contractor read before it goes out... it catches what rereading never will, for the cost of one extra prompt [here is how to set it up in any agent workflow in one prompt] 1. second prompt → paste the brief → read it as a contractor who has never seen the repo 2. list every question they'd have to ask → mark anything nobody can answer as MISSING 3. ready to go Full breakdown below.frozen5d ago3.2K57139042.59%
I ran two of my own tools together for the first time and the first session found four bugs in them. one run per condition. Sonnet 5.5. a harmless edit to one file in a real repo. the test picker printed nothing. reached through a junction, it compared the real path with the link path, decided it was not the main script, and exited silently. the same check was in five of my tools. the receipt hook watched only the Bash tool. the agent ran the tests through PowerShell, so no receipt existed. the Stop hook has an exemption for messages that say the tests were not run. it matched "I did not run the full suite" inside a message claiming 243 tests passed, and waved the whole thing through. each tool had its own tests. all green. none of them had ever run beside the other. The rule: install your tools together before you ship either one. the second session, after the fixes, is the one worth reading.frozen5d ago3.2K501010002.18%
Visa moved stablecoin prefunding into production this week! Visa now offers stablecoin payments across 180+ countries. Inyo is the first partner. The 180 countries is Inyo's payout reach, not a new Visa product in 180 countries. This is definently bullish news in my opinion. https://t.co/naOd4YQKTnfrozen4d ago3.2K491110102.22%
Checked. "All tests pass" from an agent is a claim, not evidence. Paste this into your agent: "Before you say tests pass, show me the exact test command, its exit code and the pass/fail counts from its last run. If you edited any file after that run, run it again first." Ten seconds. It can't summarize past a failing run when it has to quote the line.frozen16d ago3.2K451010042.16%
Anthropic's advisor doc has 4 things the viral Claude Code advisor posts skip. Turn it on for one session: claude --advisor opus Or inside a session: /advisor opus Turn it off: /advisor off 1. It's experimental and Anthropic API only: not on Bedrock, Google Cloud or Foundry. 2. The advisor must rank at or above your main model. Sonnet 5.5 accepts Fable, Opus 5 or later, or Sonnet 5.5. 3. Claude decides when to call it. You can't force it, but you can ask: "consult the advisor before you continue". 4. Every call re-reads the whole conversation, uncached, and counts toward your plan's usage limits. "claude --help" doesn't list "--advisor" just so you know.frozen8d ago3.2K60624022.89%
One bot per job. if you have more than five, two of your jobs are the same job. the picks from the official 54, by who you are: https://t.co/eS86H2pDAv1d ago3.2K491011052.37%
The Grok Bot team put 54 of their own bots in the marketplace. Most people will add twelve and use one. Every bot asks for something: Slack, Notion, Gmail, a spend tool. Connect half, give them a vague first task, and two weeks later ten of them haven't run since the day you added them. the marketplace isn't the hard part. choosing is. wrote the full setup: how to pick five of the 54, what to read on a bot's page before you press Add, the approvals to set before any routine runs, and the one-week test that decides which bots stay. 10 steps, 7 copy-paste prompts2d ago3.2K47910052.24%
sh*t... this is the wildest distributed inference project on the internet A from scratch webgpu engine, a webrtc mesh across every device in the room, speculative decoding that beats native llama.cpp on the same gpu, all running in a browser tab open a url, pick a model... you're live. Save this before your friends find out their phone can help run a 27b model.frozen29d ago3.1K49914112.36%
🚨BREAKING! "An unauthorised third party may have accessed encrypted customer environment variables and, in some cases, obtained the decrypted contents." That's Railway describing its own breach. • The incident: Railway disclosed unauthorised access to the systems storing customer environment variables • What's exposed: Encrypted variables were accessed, and some had their decrypted contents obtained • What's inside those variables: API keys, database credentials, and private registry secrets • Platform response: Railway paused all deployments and notified upstream providers so they can revoke keys • The attacker's goal: Railway believes it's part of a series of attacks on platform companies to compromise AI inference providers The best thing for you to do if you’re affected is to rotate all keys immediately! Source: https://t.co/8zApsvEHIQ15h ago3.1K441012012.14%
Four things Meta's official Ads MCP did that no doc told me, ranked by how badly each would have hurt in production. All four found in one afternoon on a one-dollar ad set that never ran. All four fixed, tested, and written up. https://t.co/I7oELpD3ms https://t.co/VzKdrQoIcXfrozen6d ago3.1K531012022.46%
We got flying cars before GTA 6! Thank you @elonmusk https://t.co/ZF9l5Vf3KHfrozen7d ago3.1K4975001.96%
Karpathy's post on understanding LLM output has 80K bookmarks. The first tip is a writing standard from aerospace maintenance manuals: ASD-STE100. He says to ask your LLM to explain things in it. He often finds the result a lot more readable. Sometimes he asks for "80% of the way", because the full spec is strict. I tried it on the opening of my agency article. The prompt: "Rewrite this in ASD-STE100, 80% of the way." The paragraph, before: 5 sentences, longest 28 words. After: 10 sentences, longest 14 words. One sentence, before: "You paste the brief, explain the brand again, ask for the landing page copy, get something that sounds like every other landing page, and rewrite half of it." After: "You paste the brief. You explain the brand again. You ask for the landing page copy. The copy sounds like all other landing pages, so you rewrite half of it." Same facts. Shorter sentences.frozen15d ago3.1K47912042.32%
Your agent bill isn't paying for the work. it's paying for the yeses. Jev beats Grok 4.7 and your frontier model on every yes or no in the loop: $0.00013 a call against $0.0069 and $0.0375 replace whatever makes your routing calls with Jev and keep Grok 4.7 on the writing [it takes one hook to set up:] 1/ put SelectModel in front of the agent you already run 2/ make Jev the router and Grok 4.7 the worker 3/ that's it. Full breakdown below.frozen28d ago3.1K43117012.01%
I built a tool to turn review comments into rules, then benchmarked whether review comments repeat. Mostly, they don't. The test: take a year of human review comments on vitest. Mine themes from the first 70% by date. Check what share of the last 30% land in a theme that already existed. 2.7%. With the dates shuffled, 5.2%. The ordering isn't doing anything. Review comments are overwhelmingly about the code in front of the reviewer, once. Here's the part that kept the project alive. Of the four themes mined from the first 70%, three kept recurring in the last 30%. "Failure without the fix" showed up three more times. On fastify, "add tests" came back four more times. So the honest claim is small: a handful of real rules per repo per year, not a summary of review. I put both numbers in the README, the 2.7% and the 3 of 4, because a retro tool that oversells is the exact thing it's supposed to prevent. Also in the README: one known miss. The fastify maintainer said "use describe, avoid await t.test" three times. grudge links two and stays silent under three. https://t.co/qkgxHY5Zs2frozen9d ago3.1K391010011.95%
0 to 6,000 videos a month. 150M views in 3 months. Postiz is the tool that makes running that many accounts possible. > UGC hiring: 60+ creators found through Sideshift, close to $100K spent in 3 months. > AI Replicas: Gemini analyzes creator movement, then Arcads and Seedance rebuild the videos at 10x lower cost. > Zack D Films Reposts: A viral hook, a snap cut into content that already worked, and 10M+ views. > AI Stories: Claude writes the story, Arcads generates it, and it converts. > Postiz: Posts to every platform at once, and its agent schedules everything. > The Morning Routine: Connect Postiz to your agent once, then just check the calendar. The article below shows you the full workflow. https://t.co/K3863YGoYQfrozen10d ago3.1K4796022.08%
Five things I got wrong building shortlist, ranked by how much they'd have cost you. 1. The benchmark that couldn't fail. My break-it runs on hono came back 0 missed, every time. Because npm install had written a package-lock.json, shortlist saw a lockfile change and correctly ran the full suite on every run. 0 missed, and meaningless. Now every run prints its fallback reason. 2. The JSX runtime. Nothing imports it; the compiler does. No import, no edge, missed tests. Now read from tsconfig, the Vitest config and per-file pragmas. 3. from '../../'. A trailing slash resolved to src//index.ts, which doesn't exist. One character, 31 missed test files between this and #2. 4. date-fns names its tests tests.ts. My glob found 7 of 264 test files and called the rest source. 5. Publishing. The Marketplace rejected the name "shortlist" (a GitHub user already has it) and then my 146-character description (limit 125). Now it's shortlist-tests. Rule: if your safety number is perfect on the first run, check the test before you celebrate it.frozen12d ago3.1K44106011.99%
These are the most traded crypto tokens over the last 24 hours: @GRXchain - +2428% @QuackAI_AI - +122% @AVAFoundation - +39% @BitwayOfficial - +36% @ICN_Protocol - +33% @anvil_xyz - +24% @NEXST_AI - +16% @TheSandboxGame - +15% @gitcoin- +14% Which of these do you have a bag in? https://t.co/RX9cMaPV3gfrozen4d ago3.1K44910002.05%
Karpathy explains that most people who have used an LLM are looking at a derpy chatbot! The post puts 1-2B people in that tier. to them it is a better Google search or a writing aid. Karpathy's non-tech friends say they have not much use for it, and he calls that reasonable. The surprise is who else sits there: many professionals outside math and code. execs and others spend their time talking to other people. they understand AI intellectually. it stays second-hand and abstract. The other reason: a question with a paragraph answer barely stresses the system. impact scales with ambition, problem size and horizon. No big problem, no big result. I believe just knowing who Andrej Karpathy is puts you in the 1% of AI users.frozen11d ago3.1K42106031.99%
Karpathy's one-line rule for where AI gets good fast, from his Sequoia Ascent recap: Traditional software automates what you can specify. LLMs and reinforcement learning automate what you can verify. then he turns it into a formula: capability spike ~= verifiability x training attention x data coverage x economic value his chess example: when GPT-4 got better at chess, he says it may also have been much more chess data in the training mix. not intelligence rising everywhere. so models spike where the work can be scored and the labs pointed their effort. his startup advice picks a domain that scores high on two of the four terms and low on a third. Which two, and which one is low?frozen16d ago3.1K521011002.39%
Ouch. $75k is worst case scenario tho. I think $BTC can bounce back from $80k and if not brace for impact! https://t.co/sX7yjwatCIfrozen13d ago3.1K45105022.03%
Just hit 8B tokens on Claude Code. Are these rookie numbers or is Dune the rookie?? https://t.co/areGZoXKmWfrozen12d ago3.1K49136002.23%
A bot from the Grok Bot marketplace comes with someone else's schedule. Nightly Audit Engineer, one of the Grok Bot team's own 54, runs at 4am by default. at 4am you're asleep. so the first time it needs your yes, the approval waits about 10 minutes, shows Expired, and the work never happens. you wake up to a bot that "ran" and did nothing. so before its first routine runs, set the split once: allowed in advance: reading, drafts, reports, draft pull requests. asks first: sending, posting, spending, signing, merging. then read the history after its first run. an Expired approval means it needed something you left on Ask first. Allow it, or make the routine draft instead of do.19h ago3.0K501418062.90%
Zero. That's how many failing tests shortlist skipped when I tried to make it miss. The test: take a real repo, break one source file on purpose, run the entire suite, and check that every test file that failed was on the shortlist. 48 broken files across hono and TanStack query-core. 698 test files really failed. All 698 were selected. Now the part that matters. The first honest run missed 31. Two real bugs. An import written with a trailing slash (from '../../') resolved to nothing. And the JSX runtime: the compiler injects it, so no source file imports it, so no edge existed. Break the runtime and the tests that render JSX failed without ever being selected. Both fixed, both covered by tests, and the 31 is still in the README. Speed is the pitch. "Can it skip a test that would have failed?" is the question. A benchmark that couldn't fail wouldn't answer it.frozen15d ago3.0K48910002.21%
1846 vs 1844. That is the gap between Opus 5.5 and Sonnet 5.5 on professional work across 44 occupations, and you pay $4 and $20 for one and $2 and $10 for the other. This is how you reduce cost: > Write the agency's operating file: one page that tells every model who it works for > Scope every job with Opus before anyone writes a word: deliverables, acceptance checks, and what is out of scope > Set effort before you hand anything to Sonnet: the cheap seats are only cheap when the setting is chosen on purpose > Produce the copy in volume: three versions of a hero section, each with a different angle > Run every deliverable past the Opus reviewer: every number, date and claim traced back to a source Save it, then run one client job through the five seats this weekend after reading the article below.frozen18d ago3.0K41910042.12%
You did not install an assistant. You rented a computer with someone in it. That is what Meta Muse is. Your own cloud machine, an agent living on it, a second agent guarding the only door. It keeps working after you close the app. It stops for two things: sending email and spending money. Most people asked it a question and left. The manual for using it the way it was built: - the 30-minute setup, accounts connected in order of blast radius - one first task that shows you how it fails before it can cost you anything - five jobs to hand it, ranked by how many times each hits the approval wall - what shipped on Sep 23 and what is still a date Stop asking it things. Start leaving it things.frozen26d ago3.0K48125012.19%
Everyone installed Muse and asked it a question. Nobody handed it a job. The Meta Muse Operating Manual: 3 facts, 4 payloads, 5 rungs, 30 minutes. The three facts: it runs on its own cloud computer with a guard agent approving every action, so it keeps working when your app is closed. Two things always stop for you: sending email and spending money. And you are buying tokens, not features: 100M a week free, 500M for $20, 3B for $100. The four payloads: the setup order (calendar, then Gmail as drafts only, then one booking source, payments off for a week), the First Task (five links into a spreadsheet, then one revision), the five-line task brief, and the Delegation Ladder. The ladder: research doc, grocery list from saved recipes, a table booking, a bill negotiation with a stop rule, a checkout on a one-time card. In that order. Each rung hits the approval wall more than the last. What Connect actually shipped versus what is "soon": the avatar, Mac control, the Muse email address, glasses and the Charm are all announced, not live. Article's below. Save it before you connect anything.frozen27d ago3.0K47106022.18%
A vitest maintainer left the same review comment on three pull requests this summer: "ask your agent to read AGENTS.md" The rule existed. It was in the file. The PRs kept arriving as if it wasn't. That's the state of agent rules in 2026: the real ones live in review comments, in chat, in the thing you typed for the third time this week. The instruction file has whatever you remembered to write down. So I built grudge. It reads four histories and counts what keeps coming back: PR review comments. Your Claude Code sessions. Fix-up commits. CI red streaks on main. Then it proposes the rules, each with the PR numbers or session ids behind it, and your agent asks which to keep before writing them into AGENTS.md. On fastify it found "add tests" asked six times in six PRs. On vitest, "show the failure without the fix" on seven. On my own machine, a correction I'd typed twice, which I later wrote down by hand. Paste into your agent: Install the grudge skill from https://t.co/qkgxHY6xhA, refer to the repo's AGENTS.md. Then say "retro". MIT, one file, zero dependencies.frozen14d ago3.0K491117022.66%
China just open sourced an OCR model that eats 100 page PDFs in one pass. 3B parameters, runs locally. The model is Unlimited OCR. here's how LlamaIndex's Jerry Liu scored it on ParseBench. About 46.2%. ranked 115th. it's slightly better than PaddleOCR on tables and slightly worse on layout. good at reading order. it ignores visual elements, formatting and the most complex tables. his verdict: a low-parameter model that processes any document extremely quickly, and a good milestone for that frontier. ParseBench has over 130 models on it. 115th of 130+ is a different story from "eats 100 page PDFs." Which one are you sending your documents to? https://t.co/fkTPDlID9Ufrozen7d ago3.0K52511042.43%
"License and registration!!" https://t.co/lZHbNNV2gW15h ago3.0K44911002.17%
An agent's competitor report looks finished long before it's right. three mistakes show up in this kind of job: a price pulled from the "was" field instead of the current one. a sale read as a permanent cut. a page that failed to load, reported as out of stock. Every one of them reads perfectly in the report. the client finds out when they act on it. so in month one, nothing reaches the client until you've read it. the agent drafts. you make it a product. there's a short checklist, about 5 minutes a report, and a rule for when you're allowed to trim it. You never get to trim it to nothing.frozen10d ago3.0K47811012.27%
Now that the 5.5 family is complete with Haiku 5.5, when are we getting the agent that will run my company?! Pairing whatever Anthropic puts out with Grok bot to run my life on autopilot. https://t.co/4xSvwI0AN7frozen15d ago3.0K4197001.93%
A file with two owners is a merge conflict you've scheduled in advance! That's how most agent teams break. not bad code. two workers building against two different guesses about the same interface, then editing the same routes file. the fix happens before anyone writes a line, in one charter every agent reads: > interfaces first. every signature and data shape more than one worker touches, written out exactly. workers build against them and can't change them. > ownership second. every file that will change gets one owner. shared files go to the lead. > then one last pass the charter prompt forces: list every file two roles would both need, and resolve each one. Most teams skip that pass. it's where the afternoon goes. full breakdown below.frozen8d ago2.9K44106052.20%
Everyone opens a fresh chat for every client task... this is the most complete one person agency setup that changes that. Five seats, ten steps, two models, all set up once in a weekend Write the operating file first... you're live. Save this before your next client asks for a change that was never in scopefrozen16d ago2.9K461011002.28%
89. Out of the newest 100 review comments on zod, 89 were written by bots. TanStack/query: 88. fastify over a full year: 123 of 389. I only know because grudge had to drop them before counting. A bot leaves the same comment shape on every PR, so a tool that looks for repeated review comments finds the bot first and nothing else. What's left after the drop is the interesting part. fastify: 266 human comments in a year, 125 PRs. Of those, exactly one theme a maintainer kept writing: add tests. Six PRs. Everything else was said once. That ratio is the whole argument for writing rules down. Humans now say the convention a handful of times a year and let the bot repeat the rest. The handful is what belongs in AGENTS.md, and it's small enough to find. https://t.co/qkgxHY6xhAfrozen8d ago2.9K52108022.45%
In one real Claude Code session today, the agent wrote 59K tokens and had to re-send 7.4M. 60 turns. 7 messages from me. 77% of the cost was the whole conversation being re-sent on every turn. Only 23% was the model actually writing. Five rules cut most of it: 1. Grep for the symbol, then read 40 to 80 lines around it. Never the whole file. 2. Never re-read a file that is already in context. 3. Batch independent tool calls into one turn. Every extra turn re-sends everything. 4. Run the one test, not the whole suite. 5. Past about 100K of context, write a 5-line handoff and start fresh. 11 of my 60 turns ran with 150K+ tokens of context. That is where the money went. The meter that produced these numbers reads the transcript Claude Code already writes. Zero dependencies. https://t.co/NNM5DcFoQ5frozen21d ago2.9K601224193.60%
Most people make the same decision mistake three times, because nothing remembers the first two. The /decide command in this setup writes a memo from your own notes: the options, what your notes say for and against each one, what you decided last time and how it went, a recommendation, and the one fact that would flip it. Then it leaves one section empty at the bottom: Outcome. You fill it in months later, when you know how it went. The next /decide reads it. That empty section is how a notes folder starts learning from its own track record.frozen3d ago2.9K4611110102.66%
A client sends a messy email. you do the work. then they ask for a change that was never in scope, and you make it, because nothing written down says otherwise. That is how a $2,000 job turns into a $6,000 job you were paid $2,000 for. In this setup the first seat is Opus, and it works before anyone writes a word. it takes the request as sent and returns the deliverables, what done looks like for each one, what is out of scope, and the questions you have to ask the client. Opus 5.5 is much less likely to state a wrong figure and better at noticing the one detail that does not fit. scoping is judgment. One rule comes after the scope, and every later step depends on it. This article has the full breakdown.frozen17d ago2.9K41812022.16%
Most "agent teams" are one agent doing five jobs in one context. by hour two it has forgotten the interface it agreed to, rewritten a function the tests needed, and told you it's done. The fix isn't a bigger prompt. it's a team with rules: > Opus 5.5 leads and never writes feature code. > Sonnet 5.5 workers each own their files and nothing else. > An Opus verifier re-runs every check instead of trusting the summary. the typing bills at $10 per million output tokens instead of $20. i wrote the full setup: 10 steps and 12 prompts. none of them is a "you are a helpful reviewer" prompt. each one fixes a specific way agent teams fail.frozen10d ago2.9K48109032.40%
This changes how i'd build every agent from now on... this is the most underrated agent setup on the internet right now Grok 4.7 on the writing, Jev in four seats routing, judging and gating, plain code holding the walls, all in one loop one hook in front of the agent you already run... you're live. Save this before your next agent bill landsfrozen28d ago2.9K421010012.17%
My Claude Code /usage shows 3 limit bars. The status line hands tools only 2. I found out building a free limit meter. I checked a live payload: five_hour and seven_day, nothing else. My bars right now: Session: 4% Week, all models: 18% Week, Fable: 17% The first two reach any meter built on the official data. The Fable bar doesn't, so no tool built that way can show it. If you hit a model-specific cap before the all-models one, that is the bar to watch, and it only lives in /usage. Check yours: run /usage, then compare with what your status line shows.frozen20d ago2.9K47413022.27%
Claude generated this image. Well not exactly, I created an MCP connection that uses gpt-image-2 through CLI to generate this. I simply did this cause I did not want to pay for separate API cost. What would be the likely problems I run into with images here compared to using it directly in ChatGPT?frozen27d ago2.9K51118012.46%
Claude Code cloud sessions are out of preview. close the laptop. the agent keeps working. if you already pay: $100 on Pro, $250 on Max, one time, claim by Oct 7. run /claim-credit in the CLI, connect GitHub, done. The credit spends first, then your plan. The session no longer ends when you do which is exactly what I need.frozen27d ago2.9K4367001.96%
$2,175 of tokens ran at max effort for twelve straight hours. The guide says to reserve max for measured gains only. Medium effort on the storyboard and stills beats running everything at max, matching Opus 5's best benchmark score for a fraction of the spend I replaced max effort as the default with medium for every scene except the final render pass Save this so you do not burn through all your tokens in minutesfrozen22d ago2.9K541112042.84%
You are missing out on $20 of free platform credit... Synteq just launched cloud apps, their newest product built to make deployment this simple I claimed mine just by commenting "LAUNCH" Give your repo to Synteq and go touch grass!2d ago2.9K4396012.07%
SymbioticSec scanned 1,072 vibe-coded Supabase apps! 172 let anyone delete data without logging in. The 98% headline counts any flaw, and the most common ones were missing security headers. The ones that matter are access rules: 173 sites (16%) had a critical flaw. Ask your agent before you ship: "For every table you created, who can read a row? Who can write or delete one? Show me the policy." If the answer is "anyone", it isn't done.1d ago2.8K46918022.70%
half a cow. and a flamethrower. @synteq hackathon beats a free sticker and a pizza voucher for prizes you can replace the side project nobody sees with a shot at a DGX Spark on synteq all it takes: 1/ deploy an app on synteq 2/ qrt the announcement with proof it's live 3/ that's it.2d ago2.8K45115012.23%
POV: you found the cheapest and fastest way to build AI agents for free... https://t.co/K6pIz1NftAfrozen29d ago2.7K51911022.66%
With SaaS products your users need to achieve an outcome in 3 steps or it will fail. Not entirely saying this works in all cases, but definently most. Boot up -> One Action -> Outcome For example with Postiz you have to: Link all socials -> Draft post/article -> Post to all socials With AI on the rise, there needs to be near zero friction with your products. Too many steps and you've lost all your customers. Just my own thoughts from what I have seen.frozen24d ago2.7K46810002.33%
Picture your inbox. Something has to decide what's junk, what needs a reply, and what can wait. Jev does that part and nothing else: it looks at each item and says "this is junk, 94% sure." It doesn't write back. It just sorts. 1,700 emails for 18 cents, instantly. Sounds small until you count how many jobs are exactly this. > What it is: a model that judges and sorts, and never generates a single sentence > The confidence score: every call comes back with how sure it is, so you know when to trust it > The hidden market: application stacks, support queues, inbound leads, all read one item at a time today > The rule: find an expensive queue and put Jev at the front of it > Why it's different: a whole new way to use AI, built for decisions instead of conversation Watch the video, then read the article below. Follow @neil_xbt for more content like this! https://t.co/Qz9g7ThxYpfrozen29d ago2.7K41105012.09%
Curiosity made me build a site that asks one question: Where will this product be in a few years, at the pace AI is moving? For every company it checks four boxes: - Do they own data nobody else has? - Is it wired into how people actually work? - Can they reach customers others can't? - How much of it is just the AI doing the work? Then it says whether the next big model makes them stronger or replaces them. The results surprised me. Check well-known AI products, or your own: https://t.co/Bt0Ch0ChnAfrozen6d ago2.7K40105002.02%
> Be an agent that says the tests pass. > Never call the test tool once. > Hand the run to a Jev gate. > Let it read the history, not the summary. > Watch it catch the unsupported claim. > Route the run to a human. So why is your agent still grading its own work?frozen28d ago2.7K41212012.07%
Sonnet 5.5 is half the price of Opus 5.5 per token. so why did one setting make it cost more? at lower effort, one builder's test on the same build: Sonnet $0.96 in 10.8 minutes, Opus $4.79 in 22.7 minutes, at nearly the same quality. at max effort the order flips. Artificial Analysis has Sonnet at $7.60 a task and Opus at $5.98, because Sonnet burns far more tokens on the way to the answer. same model, different bill. the setting decides which. the defaults make it worse. on the API, Opus 5.5 starts at medium and Sonnet 5.5 starts at high, and Anthropic says Sonnet pairs with Opus best at lower effort. switch to the cheaper model without touching effort and you quietly pay for more thinking than the job needs. Which setting is each of your jobs running on right now?frozen14d ago2.7K521112052.96%
With all the Opus 5.5 clips going around, I decided to create a tool that simply takes a github repo, understands the context and generates a short trailer to explain the repo. Definently needs more flair and images but for a simple straight forward animation, it does the job. It is called repo-trailer and you can find it in the replies and use it for yourself. I appreciate any and all feedback on this!frozen25d ago2.5K441011002.64%
I share a lot of useful repos here and this tool I built will make sharing even better. This is PDoomVideo, a video director powered by Opus 5.5. You can find the repo in the replies to install for yourself: https://t.co/1N0jvte6lffrozen24d ago2.3K481011023.09%
Want your email list to grow from people who already care? Start with the ones asking for your help. Sabrina shows how a useful resource can turn a comment into a lead using automation. Setup once and never miss another opprtunity The full guide is below https://t.co/bUabWf6yHCfrozen21d ago2.2K45810002.84%
I rememeber before LLMs were this advanced (2018-2019), I would be building sites with PHP and spend countless hours trying to figure out why my code was not working. Turns out i didn't close a loop with one ";" or add a "/" where it needed to be and it drove me crazy. I thought the only sites that still use this for the backend are from Wordpress but I guess not. If you still use PHP for your backend, you deserve to get hacked.9h ago2.2K38917002.94%
How to use the Grok Bot marketplace without ending up with a sidebar of dead bots: 1. write the jobs before you open it 2. match one bot per job, five max 3. read the bot's page before you press Add 4. add it, then reinstall what didn't come with it 5. scoped access, read-only first 6. approvals before any routine runs 7. one week, one job, then a scorecard 8. rewrite its description for how you work 9. let the bots hand work to each other 10. share your best one back as a template A template is a recipe. Your copy still needs the ingredients.2d ago2.1K451012123.40%
POV: you found the most complete agent UX playbook on the internet for free https://t.co/oYEQIFVXgZfrozen24d ago1.8K5410100134.92%
This is what Opus 5.5 is capable of! Same model. the difference is what you feed it before the prompt and what you say after the first render. wrote the marketing version: the context stack, the prompt template with a banned list, the still-frame gate, six director notes, and the cost math. I got this amazing output for https://t.co/uwUlC0cnrz and you can get the same with the article belowfrozen26d ago1.6K571190135.48%
.@PawelHuryn gave five models two real repos with 105 known bugs. Sonnet 5.5 at max found the most: 55.5 on average. GPT-6 Astra: 45. Opus 5.5: 41.7. the cheaper model beat the flagship at finding bugs. the price of that: 1,330 turns, about three times what Opus used. drop Sonnet to xhigh and its turns fall by more than half. its score falls to 39. so the most thorough setting and the cheapest setting are the same model on two different dials. want every bug before a release? Sonnet at max, and budget for the turns. want a fast pass on every PR? lower effort, and accept you'll miss some. one builder's test on his repos, not a benchmark. worth knowing before you set your review agent to max.frozen20d ago1.4K551411086.40%
You are missing out on the one check every viral Opus 5.5 video skipped before it shipped Anthropic's own guide names the model's default look and the fix for it, most builders never open it i turned it into a frame by frame check that fails the render before a fake number airs Give this to your Claude Code and you have a complete result in minutesfrozen23d ago1.2K501211086.89%
🚨This is insane. The timing is wild. A government-linked wallet transferred 12,267 BTC on October 8, roughly $670 MILLION in crypto over the previous 24 hours. This is definently part of the reasons the charts are tanking. Some of it went to Coinbase prime and the selling could continue.frozen14d ago1.1K541013016.87%
Two weekends polishing an AI agent, then no buyer. Flip the order: find the paid job first, and then build. This is what I mean👇 https://t.co/PVK9Zeai0gfrozen13d ago1.1K49117056.41%
Peak gaming is back ladies and gentlemen. If im inactive, it is because of Halo 3. https://t.co/gYrR7TbYrMfrozen12d ago1.0K46106006.04%
This makes a huge difference! The rules you should not miss from the article: 1. never pay a writer to make a decision 2. ask the routing question before every step 3. unsure? route up 4. judge every tool call before it runs 5. "done" is a claim. gate it with a model that did not make it 10 more rules in the article below, check it out.frozen29d ago9714686056.69%
You are paying 40% more than you need to on a model you already run... Anthropic just shipped Opus 5.5, the first of the 5.5 family it holds fable 5.1 level performance on most tasks and takes less compute to serve, so the price came down with it It is already live to use today and the full details are in the replies. https://t.co/rXnBUXW88rfrozen1mo ago95840612006.05%
POV: you and the boys retire after using this agent. https://t.co/g5kQisqtGTfrozen13d ago9084995057.49%
The /ask command. save it as a skill and your notes start answering back: "Answer the question I write after the command using ONLY the notes in /workspace/brain. 1. Check _system/maps first, then search the whole vault. 2. Read the relevant notes fully, not just titles. 3. Answer first, in 3 to 6 plain sentences. 4. Then the evidence: each note as a [[wikilink]] with the line that supports the answer. 5. Then where my notes disagree with each other, if they do. 6. Then what my notes do NOT cover that I would need to be confident. If my notes do not answer it, say "Your notes don't cover this" and stop. Do not use memory or the web unless I ask." Step 6 turns every answer into a reading list.frozen7d ago86145106037.43%
The moment your lead agent starts writing code, you've lost your lead! Opus 5.5 earns its $4 / $20 on judgment: planning, merging, deciding. At Stripe, an engineer says it directed a dozen sessions and framed conflicts plainly. that's the job. A lead who types is a worker with a bigger bill. so the lead's rules are blunt. no feature code. edit only what the charter gives the lead. a worker's receipt is a report, not proof. when two results conflict, pick one and say why. never blend them. and one rule people break by day two: never fix an escalation by quietly widening one worker's ownership. The first time the lead edits a worker's file "just this once", the ownership map stops meaning anything.frozen8d ago79046613018.35%
Do not sleep on this... Meta gave you a free agent that books the table, fills the form and haggles your internet bill. it works with the app closed. it asks before it spends a cent. most people typed a question, read the answer, and closed it. I wrote the manual for using it the way it was built: the 30-minute setup in the right order, the one first task that proves it works, and the five jobs to hand it in order of risk. Give it the brief. close the app. come back to the result.frozen27d ago73240137008.20%