Alpha Batcher

> built the app with Claude > shipped to App Store > told X it was live > got downloads > opened the dashboard > nothing > checked the code > StoreKit looked fine > checked App Store Connect > Paid Apps Agreement not active > banking missing > tax forms unfinished > Small Business Program never applied for > Apple still taking 30% > first payout still 45 days after fiscal month close > external payment copy sitting in one forgotten settings screen > reviewer would have found it instantly The app was not broken, the business setup was before you launch an iOS app, do this: > set up banking > complete tax forms > sign Paid Apps Agreement > apply for Small Business Program if under $1M > create real product IDs > test sandbox purchase > test restore > test cancellation > test subscription grace period > remove every sloppy external payment reference > then submit vibe coding gets you the app App Store Connect gets you paid

His first paying client took 4 months of outreach and paid $400 Claude later built the demo website in 8 minutes 41 seconds The 4-hour course connects those 2 numbers: > start with small client work > turn the finished steps into a reusable Claude skill > sell the same process again > price the monthly system around a 5x client return A solo agency gets faster when its skills folder stores each tested delivery process Watch the full course, then save the article below for the outreach scripts, build sequence, and pricing math

@Blum_OG right built skill saves you tokens and does the job better

one evolved skill cut Claude token use by 62% in a Kafka test the result comes from a 382-task study of procedural memory in AI agents when researchers rewrote the skills once, agents fully completed about 5 additional tasks per 100 attempts the one-person Claude business course builds the company around that same kind of skills folder this paper shows how to maintain the folder as the work changes it turns build once, skill-ify forever into a versioned loop: 1. collect traces from several jobs and models 2. diagnose the failures that repeat 3. revise the skill, with the old version preserved 4. test it on held-out tasks and other roles 5. promote the new version or roll it back one rule to steal: a skill earns promotion after it survives a new task, another role, a different model, and a changed workflow open the paper below for the full loop and first-page transfer map

@DevWith_Atul a good breakdown on what's happening under the hood

@dr_cintas GPT-5.6 SOL so huge ai model

@rohit4verse i should to try these 5 promts asap

5 PROMPTS THAT CAN TURN YOUR VIBE CODED APPLICATION INTO A READY-TO-SHIP PRODUCT: Vibe coding can get you a working demo in a weekend, but the distance between a demo and a real product is these five audits. Run all of them tonight and review every diff before you accept it. 1/ Rate limiting One script pointed at your signup route means a database full of junk accounts and an API bill you never budgeted for. Prompt: "Audit every route in this codebase for abuse potential. For each public endpoint, tell me what one request costs me (DB writes, LLM tokens, emails, storage) and what happens if a single IP sends 10,000 requests in a minute. Then implement rate limiting: per-IP and per-user sliding-window limits using [Redis / in-memory store], strictest on auth, signup, password reset, and anything that triggers an email or an LLM call. Return 429 with a Retry-After header. Finish with a table of every route, its new limit, and why you chose it." 2/ Input validation A search bar that passes raw text to your database is an open door. So is every other input in the app. Prompt: "Treat every piece of user input in this codebase as hostile. Trace each one from entry to use: query params, body fields, headers, cookies, URL segments, webhook payloads. Flag anything that reaches a DB query, file path, shell command, HTML render, or redirect without strict validation. Add schema validation at every boundary with [Zod / Pydantic]: whitelist expected types, lengths, and formats, reject everything else with a 400. Parameterize every query. Show me before/after diffs for the three riskiest paths you found." 3/ Secrets and dependencies The API key you hardcoded in week one is still in git history. Scanners find committed keys in minutes, and one package with a known CVE does the rest. Prompt: "Hunt every secret in this repo: hardcoded API keys, tokens, connection strings, credentials in logs, .env files tracked in git, keys bundled into client-side code. List each with file and line, move them to environment variables, and tell me which keys to rotate because git history already has them. Then run [npm audit / pip-audit], list every dependency with a known CVE or sitting 2+ major versions behind, and upgrade what won't break the build. Output one table: package, current version, latest, risk, action taken." 4/ Error handling and information leakage A raw 500 hands strangers your stack trace and your table names. That is free documentation for anyone probing the app. Prompt: "Break this app on purpose. Find every path where a raw error can reach the user: stack traces in API responses, database error strings, framework debug pages, unhandled promise rejections. Replace them with one global error handler: generic message to the client, full detail logged server-side with a request ID. Then hunt leakage: internal paths in responses, verbose headers like X-Powered-By, endpoints returning whole objects when the client needs three fields, user data in URLs. List every leak you closed, with file and line." 5/ File upload safety An avatar uploader that trusts file extensions will happily store a PHP shell named photo.jpg. Prompt: "Audit every file upload in this app as if the uploader wants a shell on my server. Verify file type by magic bytes, never by extension or client MIME type. Enforce a hard size limit before the file touches memory. Regenerate filenames server-side, kill path traversal, and store files outside the web root or in [S3 / object storage] with private access and signed URLs. Check whether an uploaded SVG or HTML file can execute script when viewed. Fix every gap, then write a test that uploads a PHP shell renamed to photo.jpg and prove it gets rejected." Run order matters. Start with 3, because a leaked key is damage that is already live. Then 2, then the rest. Every prompt fits under 1,500 characters, and the brackets adjust to your stack. None of this adds a feature. It decides whether the app survives contact with strangers. To do an even serious audit, check the repo in comment.

@eng_khairallah1 + more knowedges about AI for me

this is f*cking gold Andrej Karpathy joined Anthropic five weeks ago. A friend on his team just showed me the exact Claude.md file he actually uses. I dropped it into my setup. The very first response was different. Not slightly different. Completely different. Claude stopped giving generic answers and started working exactly the way I think. Bookmark it before it gets lost in your feed. Read it now, then check the article below.

@Saboo_Shubham_ never heard about this skill appreciate for sharing

/last30days is INSANELY useful agent skill. Your agent can now reads X, Reddit, YouTube, TikTok, Instagram, arXiv, Hacker News, and Polymarket in ONE command. pro tip: use this open-source skill with Fable 5 and GPT-5.6 to make it your own.

@_heyrico clearly design as always

@omarsar0 I need to working more with Frontier models

Another big reason to use combination of frontier models. Chain-of-thought monitoring is treated as a reliable safety layer for agents. This DeepMind-affiliated study shows the layer can be argued out of doing its job. Giving the monitor access to the agent reasoning trace raised approval of harmful actions by 9.5 percent on average, because the scratchpad becomes an extra channel for persuasion. The fix was model diversity. Pairing a Claude 3.7 Sonnet monitor with a GPT-4.1 fact-checker from a different family cut policy-violating approvals by up to 45 percent, versus only 6 percent when one model played both roles. If your oversight rests on one model reading another model reasoning, an adversarial agent can talk its way past it. Cross-family fact-checking is the cheaper robustness lever here. Paper: arxiv.org/abs/2607.08066 Learn to build effective AI agents in our academy: academy.dair.ai

@omarsar0 I think Fable 5 never remove

I don't think Anthropic realizes how disruptive these changes are to users. I appreciate the extension, but please stop playing games. Either keep it under the subscriptions or put it under the API already.

@Av1dlive the best guide about Fable 5

this is f**king insane Fable 5 goes away in 24 hours from Claude Code and someone just figured out how to get fable 5 reasoning in GPT 5.6 Sol with one prompt you have less than 24 hours to set this up as this takes away 20% of your usage limit here is how: 1) Paste the prompt and ask it to "make an operating manual" 2) Go the Codex App > Settings > Personalisation > Custom Instructions 3) Use GPT 5.5 / GPT 5.6 and now you have Fable 5 reasoning at 1/3 th the price save and bookmark it no matter what

@IBuzovskyi Hermes the best ai agent without any competition

HERMES AGENT HAS 3 BUILT-IN SYSTEMS THAT FORCE YOU TO PLAN BEFORE THE AGENT BUILDS. MOST PEOPLE SKIP ALL THREE AND WONDER WHY THE OUTPUT IS GENERIC. the reason AI output comes out wrong is the agent guessing instead of knowing. these three systems eliminate the guessing. 1. CONTEXT STACK (align before anything runs) three files load before your first message. the agent reads them before it reads you. SOUL.md: WHO the agent is. tone, rules, focus areas, hard limits. "you write short direct sentences. you never send emails without approval. you verify facts before stating them." user.md: WHO you are. your stack, your preferences, your timezone. auto-updated after every session. review monthly. delete what's outdated. .hermes.md: WHAT the project is. architecture, conventions, file structure. drop it in any project directory. agent discovers it on session start. without these: the agent assumes everything. with these: the agent knows the project, knows you, and knows itself before you type a single prompt. this is your alignment layer. configure it once. it runs forever. 2. /GOAL + COMPLETION CONTRACTS (spec before build) /goal is not a fancy prompt. it is a spec the agent builds against. /goal build a content pipeline that monitors X for AI news every 2 hours, drafts posts in my voice, and sends drafts to Telegram for approval. done when: - cron job runs every 2 hours without errors - drafts match SOUL.md voice rules - Telegram delivery works with approval flow the agent runs a loop. after every turn, a judge model checks: does the current state match "done when"? not "does the agent think it's done." does the EVIDENCE show it's done. completion contracts (v0.18.0): state what "done" looks like. the judge evaluates against real checks. test results. build outputs. verification scripts. not against the model's assertion. pre_verify hook: wire in custom checks. the difference between "I think I fixed it" and "the tests pass. here is proof." this is your spec-first building. nothing ships without passing the contract. 3. KANBAN DECOMPOSE (multi-lens review) one goal. decomposed across specialized profiles. each profile is a different lens on the same work. hermes kanban create "launch new feature" --triage hermes kanban decompose [card-id] the decomposer breaks the goal into tasks. routes each to the right profile by description. researcher: validates the market need. coder: builds and tests the implementation. writer: creates the announcement and docs. analyst: checks the metrics post-launch. each profile has its own model, memory, and SOUL.md. each catches what the others miss. blocked tasks escalate to you for human judgment. the board survives restarts. work resumes where it stopped. this is your multi-lens review. nothing reaches "done" without passing through every relevant perspective. HOW TO RUN THEM IN SEQUENCE: STEP 1: CONTEXT STACK (before you start) write SOUL.md, user.md, .hermes.md. the agent is aligned to you and your project. STEP 2: /GOAL WITH CONTRACT (when you build) define "done" with evidence, not vibes. the agent builds until the contract passes. STEP 3: KANBAN DECOMPOSE (when work is complex) break it across profiles. each specialist handles their piece. human gates on critical decisions. alignment → specification → multi-lens execution. the same sequence runs a product launch, a content campaign, or a codebase refactor. skip step 1: the agent guesses who you are. skip step 2: the agent decides when it's done. skip step 3: one model misses what another catches. run all three: the output stops being generic.

@Miles_Brundage didn’t using /goal a long time

Is there still much value add to /goal when Ultra is a thing?

A farmer in Hokkaido used GPT-5.6 to choose the parts, wire a Raspberry Pi, install a motor, and automate his greenhouse doors Hiroki followed its parts list and wiring instructions, and the doors worked A mathematician gave GPT-5.6 a problem he'd worked on for 3 years GPT-5.6 split the computation across multiple agents and helped disprove the conjecture Give ultra one job and define the result you can verify Greenhouse build at 0:49, parallel math at 2:29

@Haleeeemahh to be honest, i'm happy we can using Fable 5 more and more

oh wow I didn’t see this coming honestly can anthropic just keep fable 5 open to all paid plans forever?

@vikktorrrre great tips from Veee as always

you don't write because you know things you write to find out what you actually think → the fastest way to learn anything is to try explaining it in a tweet → if you can't, you don't understand it yet half my best tweets didn't exist until i started writing everything in content starts with writing

@vikktorrrre so happy right now i wanna build something with Fable 5 now

just make it part of the regular paid plan already

@Akasheth_ one more week for Fable 5 btw i'm so happy right now

Claude Fable 5 woke up and chose violence. 💀

@Akasheth_ now i wanna try this Call of duty game from grok 4.5 😂

building call of duty with grok 4.5 (high) it took 2 prompts to get here. still a lot to improve.

@Akasheth_ gemini great model, that's a fact but this model need upgrade and more features

A lot of people say gemini is worst but tbh if you can't build with gemini, claude won't save you either