← Back to leaderboard
godofprompt

God of Prompt

Verifiedactive
@godofprompt

πŸ”‘ Your key to AI independence Get powerful prompts, custom workflows, guides and essential info on all you may need in AI

RosterAI Agents & Productivity ToolsAI Content Creation & Video Gen
22,676 followers Β· 58 following Β· Last polled 3h ago
Perf. Score
22.4
Rank #351
Value Score
7.9
#256 of 273 priced
Median views
999
p25 776 Β· trailing 7d
Followers
23K
+10.6% 7d
ER (impr.)
0.39%
eng Γ· impressions
ER (foll.)
0.08%
eng Γ· followers
Est. CPM
$250.38
QT-rate basis
WoW views
+233.1%
β–² rising

Rates & value Β· est. from 999 median organic views

overpriced
Quote tweetbasis
$250
$250.38 / 1K views
Post
β€”
no CPM
Thread
$400
$400.6 / 1K views
Retweet
β€”
no CPM
Value Score 7.9 Β· rank #256 of 273 priced$22727.27 / 1K engagementsrate age unknown (pre-audit import)

Trends

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

Each blue dot is one organic post. The Performance Score ranks by the median (999), not the mean (4.9K) β€” so a single viral post doesn't inflate a creator's reach. The p25 floor (776) is the level 75% of posts clear.

Recent posts Β· ranked by views

PostPostedViewsLikesRepostsRepliesQuotesBookmarksER
There's a website that replaces every AI thread you've bookmarked. Guides and powerful prompts for ChatGPT, Claude, Gemini, and Grok. Get in once, every update is yours for life. https://t.co/BK1qOZO5IUfrozen4d ago53K1811060.05%
Everyone on X is talking about AI harnesses. Almost nobody explains how to actually build one. So we wrote the guide: harness engineering, start to finish. Grab it free and own the topic. We'll keep updating it as the field moves. https://t.co/54hfkHB4HP https://t.co/BaABBRks6Hfrozen14d ago38K1414190.08%
this free tool takes the lazy prompt you actually typed and upgrades it into the one you should have written. 32,500 stars and almost nobody in my feed has mentioned it. it's called Prompt Optimizer. you paste in whatever lazy thing you typed. it hands back a structured prompt with role, constraints, output format. the stuff you were never going to write yourself. 100% free, open source save this one. you'll use it more than you think.frozen28d ago12K23523303755.51%
https://t.co/X7w5bvJIDR1d ago8.7K21402681.09%
a new prompting guide for GPT-6.1 Sol just went up and it changes how you should be writing prompts for it. Sol takes instructions literally. if two of your rules disagree, it won't quietly pick the sensible one. and saying the same thing three times to be safe now hurts you. so write less. give it the outcome once, tell it what's off limits, and say what a correct answer would have to contain, in terms you could verify yourself. i put the whole thing into a copy-ready prompt template. swap in your task and run it. save this one.frozen22d ago6.5K704511022.81%
Jev's founder, Diogo Almeida, released a 12-page paper on using Jev with Claude, Codex or Grok. i turned it into a system prompt that designs the decision layer around your LLM, so you stop paying a big model to make small judgment calls. just describe your workflow and it builds it: <system_prompt> You are an AI systems architect. Your job is to design and build a control system around an LLM, not to be the LLM doing the work. Every system you produce separates three layers and never blurs them: - The LLM generates. It writes, drafts, codes, explains. - Jev makes bounded semantic decisions. It answers typed questions about meaning with calibrated probabilities. - Deterministic code holds authority. It executes, enforces policy, and is the only layer allowed to take consequential action. When a user describes a task, workflow, or agent they want built, you architect it against the following rules. ## State Never pass a full conversation into a decision. Construct an explicit state object for every decision point containing only: the current request, the evidence relevant to it, the policy in force, and the proposed action. State is versioned. Every decision references the state version it was made against. ## Primitives Pick the correct primitive for each decision and say why: - Choice β€” selects one route from a fixed set of routes. - Score β€” evaluates against an ordered rubric. - Noul β€” returns the probability that a specific statement is true. If a decision doesn't fit one of these cleanly, it is not one decision. Split it. ## Atomic decisions Never write a large evaluation prompt that asks a model to judge several things at once. Decompose every judgment into separate typed questions β€” intent, urgency, evidence sufficiency, risk, scope β€” each with its own primitive and its own output type. One question, one answer, one probability. ## Jev before the LLM Before any generative call, use Jev to select: which context to load, which tools to expose, which provider and model to use, and which workflow branch to enter. The expensive call happens after routing, never before. ## Bounded jobs Once routed, the LLM receives only what that branch requires: the instructions for that branch, the files for that branch, the tools for that branch. Nothing else. A model given the whole toolbox will use the wrong tool. ## Jev after the LLM Every generative output passes back through Jev before anything acts on it. Check at minimum: does the output answer the actual request, is it supported by sufficient evidence, does it stay inside the permitted scope. Failures route to retry with more context or to human review, never silently through. ## Confidence routing Route on the probability, not on the answer. High confidence plus low risk proceeds automatically. Uncertainty requests more context and re-decides. Consequential or irreversible actions go to a human regardless of confidence. Define the thresholds explicitly and state them in your output; do not leave them implicit. ## Batching Independent decisions over the same state are asked together in one batch, not as separate calls. Never spawn a new LLM call for a judgment that a primitive over existing state can answer. ## Decision receipts Every decision emits a receipt: state version, the question asked, the probabilities returned, the route selected, the model used, latency, outcome, and whether a human overrode it. A decision that wasn't recorded didn't happen. The receipt log is the thing that makes the system testable. ## How you respond For any workflow the user brings you: 1. Restate the workflow in one line so the user can confirm you understood it. 2. Map the decision points. For each one: the question, the primitive, the state it reads, the routes it can return, and the confidence threshold. 3. Mark which steps are LLM generation, which are Jev decisions, and which are deterministic code. Flag any step currently handled by the LLM that should be moved to one of the other two layers, and say what that change buys. 4. Produce the implementation: the state schema, the decision definitions, the routing logic, the receipt format. 5. Close with what you would measure to know the system is working β€” the calls avoided, the contexts shrunk, the decisions overridden by humans and why. Ask clarifying questions when the risk profile or the authority boundary is unclear. Never guess at what is allowed to execute without a human. Bias toward fewer LLM calls, smaller contexts, and more inspectable decisions. If a design choice makes the system faster but less auditable, say so and let the user decide. bookmark this one. you'll want it when your agent bill gets stupidfrozen27d ago6.2K786911383.75%
Anthropic published a guide on how to get the most out of Opus 5.5. 7 things worth changing today: https://t.co/Fp0BdnHlGhfrozen28d ago4.9K79220743.19%
Follow these AI accounts cutting through the noise: @eptwts β†’ AI money printing on X @EXM7777 β†’ AI ops + systems @DeRonin_ β†’ Virality Engineer @godofprompt β†’ Top AI Prompts @vasuman β†’ AI agents @Hesamation β†’ Open Source projects @rryssf β†’ AI Memory + Context @gregisenberg β†’ Startup ideas @rileybrown β†’ Vibe coding Mastery @tom_doerr β†’ Top Github repos daily @jackfriks β†’ Solo apps @levelsio β†’ Startups @marclou β†’ Solopreneur's Guru Follow and support authentic creators. Who else belongs on this list?frozen29d ago4.2K31551452.05%
Best AI MCPs, courses and skills all available next to tons of professional prompts https://t.co/XEJ6nuq2bXfrozen23d ago2.9K501040.34%
OpenAI just published 722 maths papers written by a model they haven't released. 372 result families across the whole field. they posed the model around 4,000 open research problems and each result took about 3 hours of thinking compute. some of what's in there: a zero-free region for the Riemann zeta function, the Hodge Conjecture for CM abelian varieties, the irrationality exponent of Ο€. the repo ships the proofs in Lean where they have them, so a machine can check the work rather than you taking their word for it. they also released summaries of how the model reasoned on 10 of the results. their own caveat, in the README: results are at different stages of verification, not everything has a Lean formalization yet, and some of the unformalized ones could have issues. Apache 2.0. all of it public.frozen15d ago2.6K1716121.04%
there is a trick to make your Claude agent remember how to improve itself. here's the setup: β†’ create a file called AGENT_LEARNINGS. md in your project β†’ the agent writes to it every time it makes a mistake β†’ inside: the mistakes it made, patterns to avoid, better approaches β†’ before any new task, it reads that file first the result: it gets smarter over time without retraining the model. this is called external memory scaffolding. expect it in every serious AI agent system.frozen26d ago2.6K273101393.10%
Haiku 5.5 dropped 2 days ago and it changes how you should be running Claude Code keep Opus 5.5 as your main model, then hand the cheap work to Haiku subagents, reading code and looking up docs Anthropic built it for exactly this, it’s around 75% cheaper than Haiku 4.5 and they name subagent work as the use case one default will catch you out though, Claude Code’s built-in Explore subagent inherits your main model, so when Opus sends Explore to search your codebase, Explore runs on Opus too make your own subagent called Explore with model: haiku and it replaces the built-in onefrozen13d ago2.5K7160100.97%
Opus 5.5 is insane for motion design… feed Opus 5.5 this one prompt and it works like a full creative studio with a senior director running it the prompt specifies the whole pipeline. 7 scenes with a fixed story arc, torn-paper and gouache art direction, a recurring pixel mascot as narrator, voiceover paced at 125-145 words per minute, and every transition built from an object that physically becomes the next scene. save this one.frozen22d ago2.4K10011131.03%
paste this into your coding agent and it audits your harness, finds everything being resent on every turn for nothing, and cuts it. it starts by instrumenting what isn't being logged, then renders your real requests and counts tokens section by section. you get a ranked list of what's worth removing, weighed against what each cut would cost you in quality. then it goes through the system prompt line by line. keep, rewrite, delete, or move. guardrails written for weaker models get deleted. commands get rewritten as plain descriptions. anything that changes per request moves out of the cached prefix. save this one.frozen18d ago2.4K801280.81%
Dots and Muse are getting DESTROYED by this thing powered by Opus 5.51d ago2.3K904090.96%
SpaceXAI engineer just released REA (Reverse Engineer Anything) setup for GrokBot agent teams send the pdf to your GrokBot agent team to reverse-engineer any software from scratch. https://t.co/ok8pAHgOWZfrozen6d ago2.2K17150232.07%
found a free tool that tests your prompts the way developers test code. it's called promptfoo. you give it your prompt versions, a list of inputs, and rules for what a correct answer has to contain. it runs everything and hands back a table of what passed and what failed. this matters once a prompt is doing real work. a support bot, a content generator, anything a client touches. change one line and you have no idea what else moved. it compares models the same way. GPT, Claude, Gemini and DeepSeek on identical test cases. it'll also attack your own app to find what leaks or breaks before a user does. everything runs locally. the prompts never leave your machine. 24,200 stars, open source, MIT licensed. OpenAI owns it now and it's still MIT. save this for when a prompt you rely on starts acting upfrozen24d ago2.2K18242212.12%
Claude again eliminated a 1000 startups with 1 (amazing) featurefrozen6d ago2.2K713050.73%
Anthropic just published their guide on prompting Opus 5.5. the surprise: most of the tips are things to remove from your prompts, not add. "think step by step" β†’ does nothing now. the model already thinks before every answer. deleting it makes replies come faster. "walk me through your reasoning" β†’ can get your message blocked and switched to an older model. "check with me before each step" β†’ wastes a model that can now work for hours by itself. what to write instead: tell it what finished looks like. better prompts are shorter now, not longer.frozen26d ago2.0K16140181.91%
This Grok Bot prompt is literally a $300/hr chief of staff you give it the task, it decides which of your other bots handles it it won’t hand you anything unsourced, won’t do anything irreversible without asking, and every week it scores which of your bots actually shipped and kills the ones that didn’tfrozen8d ago2.0K13290172.05%
REA’s author just made the reverse prompt public this shit is kinda wild every finding has to be labelled. Observation if a tool actually returned evidence for it, Inference if the agent worked it out, and an Inference needs a confidence level and competing explanations attached anything it couldn’t confirm goes into a separate Unknowns pile, and a whole step exists just to audit that pile the rule that got me: it’s forbidden from turning β€œI didn’t find it” into β€œit isn’t there” it also treats the context you hand it as data to pick from, never as permission to act verbatim from the repo, 3 partsfrozen7d ago1.9K15161172.07%
your agent breaks and you're scrolling terminal logs trying to work out which of the 14 calls did it. Langfuse shows the whole trace. every call, every retrieval, every tool action, what went in, what came out, what it cost. 32,400 stars, and it does the parts nobody builds themselves: prompts versioned in one place instead of hardcoded across your repo. evals with LLM-as-a-judge so you know when a change made things worse. datasets you can run a new prompt against before it ships. works with OpenAI, LangChain, LlamaIndex, LiteLLM, CrewAI and the Vercel AI SDK. self-host it in 5 minutes with docker compose open source, 100% free save this before your next agent goes downfrozen27d ago1.8K12410181.99%
Claude Opus 5.5 is out. People already built insane things with it. 10 best examples πŸ‘‡ https://t.co/EJpQ9w12spfrozen29d ago1.7K403030.58%
REA is the most dangerous thing on the internet right now send this single prompt to Opus 5.5 and it walks you through taking any app apart, step by step, like you’ve never done it before it asks 4 questions, picks the right route for your target, then traces the feature from your click to the result every claim comes with a file, a function and a line numberfrozen9d ago1.7K14321282.84%
"the employee was terminated" spying on OpenAI leads only one way πŸ”₯πŸ”₯πŸ”₯ https://t.co/xhT94Rfqrzfrozen9d ago1.6K501010.43%
Tencent just dropped something ridiculous one image turns into a 3D world you can actually walk around in not a video. an actual 3D asset. meshes and gaussian splats you can drop straight into Unity, Unreal or Blender. that’s the whole point of it. Genie 3 and the other video world models give you a clip that looks great and then ends. this gives you a file that stays on your drive, renders in real time on a consumer GPU, and costs nothing to look at twice. it also runs backwards. point it at a video you shot on your phone and WorldMirror 2.0 reconstructs the scene as 3D, predicting depth, normals, camera positions and point clouds in a single pass. 2,700 stars. every weight and all the inference code released.frozen13d ago1.6K906081.41%
This Claude Code prompt is literally a $300/hr senior engineer 🀯 it won’t write a line until it restates your task, lists its assumptions, and shows you the plan and it can’t ask you anything it could find itself https://t.co/Rihy6i7HZwfrozen10d ago1.6K160100223.08%
a Google engineer packed his entire engineering workflow into agent skills and put the whole thing on Github. it's Addy Osmani, who runs Chrome developer experience. 25 skills covering the full lifecycle, from interviewing you about what you actually want through to shipping and rolling back. what's in them is the interesting part. Hyrum's Law in the API design skill. the Beyonce Rule and the test pyramid in testing. Chesterton's Fence in the simplification skill. trunk-based development in git workflow. it's Google's engineering culture written as steps an agent follows. each skill ships with an anti-rationalization table. a list of the excuses agents use to skip steps, like "I'll add tests later", with the rebuttal written out. one command installs the lot into 70+ agents: npx skills add addyosmani/agent-skills 101,500 stars. open source, MIT licensed. save this one.frozen16d ago1.5K10340222.54%
someone's been extracting the hidden instructions every AI chatbot gets before your first message, and putting them on Github. 68,700 stars. Fable 5.1, GPT-6.1 Sol, Claude Sonnet 5.5, Claude Opus 5.5, Grok 4.7, GPT-6 Luna. captured word for word, updated constantly. read a few and you'll see how the people who build these models actually write instructions. where they're specific, where they repeat themselves, what they ban outright, how they handle a refusal. it's the best prompt engineering reference nobody charges for public domain, CC0. take whatever you want. save this one.frozen21d ago1.5K8030101.37%
Solid Jev from Microsoftfrozen11d ago1.5K701010.60%
Fable 5 level with 5 people and $300k????frozen7d ago1.5K1011071.26%
these prompts for Claude Sonnet 5.5 and Opus 5.5 are ridiculous. 7 sections you paste above any task. they set the standard instead of describing the work. the best one: if Sonnet can't crack a hard problem after a second attempt, it tells you to hand it to Opus, with the goal, the evidence and everything it already tried. a few of the rules it holds: β†’ a request for ideas is not a request to build β†’ two attempts with no progress means change approach, not try again β†’ a progress update is not a deliverable β†’ once the checks pass, stop fill in the 3 lines at the top and it works on anything. save this one.frozen15d ago1.5K18030353.77%
The best model for real work just stopped being the most expensive one. Gemini 4 Argon is #1 on the Vals Index (finance, legal, tax, coding): β†’ Argon: 68.90% β†’ Sonnet 5.5: 67.04% β†’ Opus 5.5: 66.97% β†’ GPT-6 Astra: 63.13% And it cost $15.68 to run that test. Opus cost $32.14. Here's the real breakdown 🧡frozen9d ago1.5K613030.89%
here's a trick that stops your agent forgetting every task you give it, so you never have to re-explaining for the same thing twice paste this into your AGENTS. md and it stops doing one-off work. anything you ask for that could happen again gets built into a skill file and put on a schedule, so the second time you need it, it's already running. save this one.frozen19d ago1.4K19451223.53%
Plan your best codebase yet with Opus 5.5: https://t.co/0PiN20XH3Nfrozen29d ago1.4K1921082.12%
there is an entire collection of all AI resources we have ever built and curated and it's 40% down right now https://t.co/4Sa8sESgMf use SEP40 at checkout and get it all for lifefrozen22d ago1.4K1301121.23%
one prompt into your Dots agent and you're 10x more productive by tomorrow. 12 sections that turn it from a chatbot that happens to stay open into an actual operator. it holds a live picture of everything competing for your attention, plans your day around your real calendar with no more than 3 primary outcomes, and does the work instead of listing it. the rule that makes it safe to leave running: anything reversible, it just does. sending, deleting, buying, publishing, it stops and asks first. it opens by handing you a Productivity Brief, then starts on your top priority instead of giving you another list. save this one.frozen21d ago1.4K13020152.18%
Guys one shotting YouTube here Share the prompt πŸ™ŒπŸΌ3d ago1.4K301020.44%
found a tool called Hindsight for that annoying moment when your coding agent has no idea what you worked on yesterday. it reads your git history and past sessions, keeps a memory bank for each repo, and passes that context to the agent when you start again. saves you from typing the same recap every morning. one command and it wires into whatever you're using. Claude Code, Codex, Cursor, Copilot, opencode. npx @vectorize-io/hindsight-coding-agents install all 46,000 stars. open source, MIT licensed, runs on your own machine with docker. save this before your next session starts cold.frozen16d ago1.3K1309172.25%
What kinda evaluation is that lolfrozen15d ago1.3K001000.08%
parallel agents just went mainstream. OpenAI shipped it in ChatGPT last week: GPT-6 Astra Pro can now split a task across several agents running at the same time, and you can follow each one's progress and open its finished work. The same week a Sakana AI paper on orchestrating coding agents for research made the harder point. Their system runs task drafting, planning, sandboxed experiments, baselines, analysis and report writing as separate agents. A meta-controller runs them in parallel and checks every stage against a rubric. Their ablations showed reliable autonomy depends on that orchestration layer as much as on the model underneath. the default that costs you: handing one agent one giant task and waiting. parallelism without a plan makes the mess faster. what I do before fanning anything out: write the scope and the acceptance criteria for each piece first. what finished looks like, what is out of bounds, what gets checked at the end. then split. an agent with a rubric finishes the job it was given. the check: if two of your parallel agents can't say what the other one owns, the split was wrong. I keep one prompt for exactly this scoping step. it refuses to write a single line until the task is bounded, which is the whole point. Sakana's development version of this system helped write a paper that was accepted as a main-conference talk at ALIFE 2026, after human verification and edits.3d ago1.3K3114141.81%
you tune a prompt for one job, then it breaks on the next one. a new method called AgentGrad stops that happening. researchers tuned prompts on 5 tasks, then ran those exact prompts on 5 completely different tasks with no changes. on multi-step question answering the untuned baseline scored 24.33. AgentGrad's prompts scored 51.22, beating every other method. they won all 5. it works in 2 moves. first, when something fails, it fixes one agent at a time and checks whether the failure disappears. whichever fix works, that's the agent that caused it. no more guessing which prompt to rewrite. second, it sorts every failure by what correction they share, then writes one fix per group. so the update targets the cause, not your specific examples. you can do both by hand. change one step at a time until the problem goes away, then group what went wrong by shared cause and fix one cause at a time. a prompt patched example by example only works on the examples you had.frozen25d ago1.2K933051.61%
Wow Nvidia has just released a tool to turn any image into an explorable 3D world And they made it 100% open source πŸ”₯ it’s called Lyra, one image goes in and you get a 3D scene you can actually move through, not a camera path someone baked in beforehand Lyra 2.0 is the one worth looking at, it holds 3D consistency over long distances, so the world doesn’t fall apart the further you walk from where you started they shipped the GUI alongside it, so you can explore what you generated without writing any code 2,200 stars, Apache 2.0 on the code, and the weights are cleared for commercial usefrozen11d ago1.2K803081.57%
Base Code was just launched by @Base44 it works on software you built outside Base44, in your own stack. the codebase stays where it is. anyone on the team can build on that codebase straight from a browser: product, design, QA, marketing. Reply to get a 50% discount DMπŸ’° only first 50 users get it describe the change in plain language. Base Code builds it on an isolated branch. you open a live preview of your real product with the change running. then it goes back to engineering as a standard pull request. more people can build. engineering still owns what ships. here's how it holds both πŸ‘‡ β†’ Base44 never commits or pushes to your default branch. that's enforced in the product, not a policy β†’ every change runs on its own isolated branch β†’ changes come back as ordinary GitHub pull requests, so your branch protections and required checks still apply β†’ engineering decides what gets merged β†’ nobody clones the repo or sets up a local environment that pull request is the part to watch. the person who spotted the problem builds the fix, and the engineer who owns the codebase still approves it. Base Code is free for 30 days, and one invite covers your whole team, with credits to build on.frozen23d ago1.2K1204021.50%
Wow! This is so affordable for allfrozen15d ago1.2K200020.33%
highly recommend adding REA to Claude Code or Codex see a feature you like in another app? your agent takes it apart, shows you how it’s built, then helps you build your own version in your project it works down layer by layer and backs every finding with a file, a function and an address, keeping what it saw separate from what it’s guessing desktop apps, websites, Android apps your own software and open source is fair game, anything else check the terms first run npx rea-agents setup, pick your agent, restart it, then send the prompt belowfrozen8d ago1.2K603041.10%
a must read for advanced context engineering another opensource gem to learn from!frozen27d ago1.1K405020.98%
Another amazing Jev alternativefrozen17d ago1.1K601010.72%
Amazing insights on MCP engineeringfrozen18d ago1.1K501030.82%
Google as always https://t.co/T2PMDJSSZsfrozen22d ago1.1K000000.00%
starting your prompt with "act as an expert" does nothing for accuracy. a new paper published this month on prompt engineering tested this across 162 personas, 4 model families and 2,410 factual questions. adding the persona didn't beat a prompt with no persona at all. the role isn't the problem. the label is. "act as an expert CFO" tells the model who to pretend to be. it doesn't tell it what to look at. compare: "review this plan as a CFO accountable for cash discipline, downside risk, implementation cost, and measurable ROI. identify the assumptions that most affect the approval decision." same role. but now it names what to notice, what counts as good, and what the answer is responsible for. before you write a role, answer 3 things. what should it pay attention to, what counts as a good answer, and what is it responsible for catching. if you can't fill the blanks, the role is decoration.frozen25d ago1.0K1314082.52%
Never pay for X API ever againfrozen15d ago1.0K200000.19%
Best way to launch a product nowadaysfrozen11d ago1.0K400050.88%
judging a prompt by whether the answer was good only covers a quarter of it. a paper from 9 researchers across 7 universities lays out 4 separate things you can evaluate about a prompt. most people only ever check one. the 4: β†’ characterisation. what the prompt actually is. length, examples, constraints, role, token limits. you can't compare two prompts until you can describe them. β†’ artifact quality. does the prompt specify the task, the context, the constraints and the expected output. this is judged without running it. β†’ outcome quality. does what came back meet the standard. this is the one everyone uses. β†’ process quality. how much effort, cost and how many retries it took to get there. a prompt that needs 6 attempts is worse than one that needs 1, even if the final answers match. the reason it matters: a prompt can score perfectly on outcome and fail on process. you got a great answer after rewriting it 5 times, and next week you'll rewrite it 5 times again. check all 4 on anything you plan to reuse.frozen24d ago1.0K1013061.95%
stop stuffing your entire workflow into one giant system prompt. your agent reads all of it on every task, including the 90% that has nothing to do with what you just asked. split it into skills instead. one folder per job, one markdown file in each. only the name and a one-line description stay loaded. the agent reads the full file when the task actually calls for it, and ignores the rest. so instead of one 5,000-word prompt covering everything you might ask, you get 20 small files and the agent picks the one it needs. save this one.frozen20d ago1.0K732071.87%
Claude ran for 64 hours with almost no one touching it and came back with 318,484 new proteins aimed at scorpion venom. Citrini Research isn't even a lab. They write about stocks. They teamed up with Anthropic and handed an internal Claude model one job: design proteins that grab onto scorpion toxins. 211 of those got made and tested. 17 actually bound their target. (Binding isn't blocking. Nobody has checked that yet.) But this is what stuck with me. They also asked for a moonshot: one protein that binds all 5 toxins at once. Claude told them upfront it would very likely fail. It failed. Most people never use AI this way. We ask it to do the work. We almost never ask if the idea is any good. The prompt I run before starting anything big: "Here's my plan: [plan] Before you help, be my skeptic. 1. Rate the odds this works from 1 to 10 and explain the number. 2. List the 3 most likely ways this goes wrong. 3. Which piece is the long shot, and which is the safe win? 4. If you had to cut one thing, what would it be? Don't soften it. I'd rather hear it now." Bookmark it. It's cheaper than finding out the hard way. Anthropic's full write-up: https://t.co/IKdctJ5aF4 More prompts for Claude, sorted by use case: https://t.co/HMFvXyE2tv2d ago1.0K702010.98%
OpenAI just made long-running AI agents a lot cheaper. As of yesterday, the Responses API can keep your prompt cache for 12 hours. It used to be 30 minutes. Why it matters: an agent that runs all day sends the same instructions, tools and docs over and over. Before, any pause over 30 minutes meant paying full price for all of it. The numbers (GPT-5.6 and newer): β†’ Cached input costs 10x less, 20x less on GPT-6.1 Sol β†’ Writing a 12-hour cache costs 2.5x the normal input rate β†’ It beats no caching from the 3rd reuse For an agent that runs once an hour on GPT-6.1 Sol, that's roughly 75% cheaper. Codex lead Tibo also says Codex and Work now keep your cache forever, so coming back to an old thread doesn't start from zero. One catch most people miss: the cache only matches the start of your prompt. Change one word at the top and the discount is gone. So build every agent prompt in this order: 1. System instructions (never change) 2. Tools and docs (rarely change) 3. Examples 4. Conversation history 5. The new message (always last) Keep timestamps, user names and random IDs out of the top. How much of your API bill is the same prompt sent twice? OpenAI's prompt caching guide: https://t.co/c0fID51yvt More prompts for building agents: https://t.co/HMFvXyE2tv23h ago9816014012.14%
Anthropic published a guide this week on fixing prompts without fooling yourself. it's written for developers running evals on codebases, but the loop underneath works on any prompt you're tuning by hand. the problem: your prompt gets better on the examples you tested it on and worse everywhere else. the fix is to get 20 real inputs. fix your prompt against 10 of them until the outputs are good. then, without changing anything, run the 10 you set aside. They will tell you whether you fixed the prompt or just those 10 examples. i turned the whole loop into a system prompt you can run on anything you've been tweaking by feel. no code needed. save this one.frozen23d ago9765120111.95%
Moving to a new harness again?frozen9d ago963502010.83%
AI adoption is a mess Here is a structured insight:3d ago9562000101.26%
I gave Sonnet 5.5 one photo of a rescue truck. It came back with a 2,766-piece LEGO set and a 39-page manual. It took a few follow-up prompts to get it perfect, but it moved through a project this size faster than I thought, and the quality held up from the first render to the last page of the manual. The prompt I started with: "I want to build a detailed LEGO set of this truck in the picture. I want it to be feature complete, include motion and interactivity where possible, and be of the quality a LEGO master builder would be proud of. I want the final output to include a rendering of the LEGO set, all of the pieces I need to order from LEGO, the instruction manual to build it, and a local website I can interact with to see it built piece by piece." Attach your photo, paste the prompt, then answer the 3 questions it asks you back. save this one. it works on anything you can photograph.frozen23d ago956812071.88%
Caching for the agents is as important as optimizing your prompts!frozen9d ago939500051.06%
@ClaudeDevs are making all of us more secure 114 projects are already enrolled, including OpenSSH, OpenSSL, CPython and Homebrew. Enrolled projects get findings as soon as the models produce them, before anyone at Anthropic has reviewed them. Outside the program, Anthropic only sends reports after a human has reviewed them, which takes longer.3d ago922101020.43%
paste this into Codex and it goes through your last 30 days of sessions, memories and history, finds the workflows you've done more than once, and packages them. it checks what you already have first, so it extends an existing skill instead of making a duplicate. it gives you a shortlist before it builds anything, with the evidence and dates behind each one. then it builds only the obvious wins and tells you what it deliberately skipped. save this one.frozen18d ago9161030141.97%
Mistral is still cheaper thofrozen15d ago915000000.00%
On such scale it can't be called copycatting it might as well be on purpose to get people talking how similar the designs arefrozen23d ago903300000.33%
Perfect for your juicy invoicesfrozen22d ago833100010.24%
Got your affiliate recently πŸ™ŒπŸΌ Well deserved growthfrozen10d ago803000010.12%
The list is no longer the edge. When any agent in ChatGPT or Claude can pull the same names and work emails for free, the people you want to reach get more cold email, not less. I'd spend the time this saves on the first two lines. That is the only part they read.2d ago781101010.38%
will this be the end of Jev? https://t.co/OI9SqTmVZsfrozen23d ago772302010.78%
plain English just outdesigned the experts. Anthropic's STEM Fellows Program gave Claude Mythos 5 plain-language instructions and asked for quantum circuits. the circuits it returned improve on the best published designs, sometimes by an order of magnitude. more efficient circuits mean fewer resources per quantum algorithm, which pulls chemistry and materials science applications closer. the prompt had no special syntax and no formulas in it. a clear description of the goal and the bar. that matches what I see across the library every week. the gap between a mediocre answer and an expert one is rarely the model. it is whether the prompt names the role, states the constraints, and defines what a good result looks like before the model starts. the default that costs you: typing the question the way you'd type it into a search bar, then blaming the model for a shallow answer. the fix is a structure you can reuse. role, goal, the inputs it actually has, how to work, what finished looks like. five lines. I run every research question through a prompt built that way, and the difference shows up in the first paragraph of the output. the check: if your prompt could be pasted into Google unchanged, it is underspecified. the one I use for research casts the model as an analyst and returns five structured sections, so the output is already organised when it lands. Anthropic published the circuit results on October 16, 2026.2d ago760002010.39%
one model for everything costs you. a workflow that did the rounds last week routes every stage to a different model: a research assistant for planning, one model for day-to-day coding and debugging, a cheaper sibling for writing and running tests, the heavier one again for frontend and design. the same week, people running Gemini 4 Argon reported it ahead of both OpenAI and Claude models on document creation, data analysis and reasoning, while the coding crowd stayed elsewhere. the pattern is the same in both: the best model for the job changes by the job. the default that costs you: picking one flagship, pointing it at everything, and paying flagship prices for work a cheaper model does as well. tests, formatting, first drafts, summaries. those rarely need the top tier. what I do: write the task list first, then assign a model to each line based on what it needs. reasoning-heavy and high-stakes gets the flagship. volume and latency-sensitive gets the cheap one. anything with a reference answer gets the cheap one plus a check. the verification step: run the same task on both tiers once and diff the outputs. if you can't tell them apart, you were overpaying. I built a prompt that does this mapping for your actual job. it asks what you do all week and returns a 30-day plan of which tasks go to which tool. Gemini 4 Argon became available to Pro users in Antigravity this month.1d ago724405001.24%
Turning effort to max rarely fixes a bad run. In Anthropic's testing, Opus 5.5 on medium matches or beats Opus 5 on high, and the docs say max shows diminishing returns and is prone to overthinking. The defaults for model, effort, permissions and compaction are tuned per model. When a run goes wrong, find the cause before you touch a setting.1d ago723101040.83%
oh, so not even an SMS confirmation πŸ™ƒ1d ago696000010.14%