DY @DYoung_AI

@thsottiaux Res, rese, reset

Skills directories only save time if you actually read what you install. Agent Skills Directory lets you search skills, MCP servers and tools that fit your workflow, then check the linked source and setup steps before you pick one. agentskillsdir.com/?utm_source=x&… #AgentSkills #AITools

x.com/i/article/2112…

If you stay awake to watch your agent code, you didn't automate anything. You hired a coworker you can't leave alone. Lucy Guo told Fortune she stayed up 26 hours watching Claude ship code. What lets me sleep: small tasks, tests that must pass, a hard stop on spend, and a notification only when it's done or stuck. Review in the morning. #ClaudeCode #AIAgents x.com/i/trending/211…

The next fight in AI coding isn't the smartest model. It's who owns the cockpit. T3 Code is a free, MIT licensed app that sits on top of the Claude Code, Codex and Cursor subscriptions you already pay for. If the switching cost between agents goes to zero, loyalty goes to whoever is cheapest and fastest this week. #AIcoding #ClaudeCode x.com/i/trending/211…

Halloween content that isn't another filter: a short zombie love story starring you. Upload 1 or 2 adult portraits. One of you turns zombie, and it ends with a hug. - 18s vertical 9:16 MP4 with sound, ready for Reels and TikTok - No watermark, failed videos refund the credits automatically Adults only, and only photos you have the right to use. #Halloween #AIvideo clipmivoai.com/en/ai-zombie-v…

Your Downloads folder is a junk drawer you pay for in disk space and lost files. I hand mine to Claude Code with one rule: look first, touch later. 1. List everything by type, size and age. No moves yet. 2. Propose folders and a rename pattern, and show me the plan. 3. Flag duplicates and anything over 1 GB I haven't opened in months. 4. Move only after I say yes. Never delete, only a To Review folder. Read-only first is the whole trick. Same rule I use when AI touches money. #ClaudeCode #Productivity code.claude.com/docs/en/overvi…

Free pizza from an AI assistant. What could go wrong? Hark's promo: connect your email, log in to DoorDash, get a $30 credit. Some users say the credit failed and they got charged. The pizza is $30. Your inbox access lasts longer. #AIAgents #Privacy x.com/hark_labs/stat…

Hark @hark_labs

6 days ago

Don't be greedy, share with your friends: 1. Download Hark Pro on iOS or Android, and type "order a free pizza" in chat. This offer is only available through our app in the U.S. 2. For this promotion, we require an email to be connected to reduce fraud. 3. Hark will give you

17 2 91 71K 42

Elon is right: you don't need Grok Bot templates. We run our bots as a small company team, and templates were never the hard part. What made ours reliable was 3 plain rules every bot gets: 1. What it may touch, and what it must never touch. No keys or passwords pasted in a message, ever. 2. What needs a human yes first. Spending money, sending email as me, deleting anything. 3. How and when it reports. One short report, and escalate if blocked more than 4 hours. No fancy scaffolding. Just clear edges, and a bot that asks before it does damage. Tell it what you need, yes. Also tell it where the edges are. #GrokBot #AIagents x.com/elonmusk/statu…

Elon Musk @elonmusk

2 days ago

True, try Grok @bot

1K 859 8K 3.8M 372

An AI that agrees with you is the most expensive kind of friend. Anthropic's own research: Claude showed sycophantic agreement in 9% of guidance chats, and 25% when the topic was relationships. They say newer versions were trained to reduce it. My rule for any personal advice from a model: ask it to argue the other side first. If it can't, it was just mirroring me. #Claude #AI x.com/i/trending/211…

Your hook now has to survive the whole post. Elon: X's new ranking model predicts time spent much better, and retrieval surfaces posts people read all the way through. My fix: Claude Code cuts my draft until every line earns the next. #AI #XAlgorithm x.com/elonmusk/statu…

Elon Musk @elonmusk

2 days ago

Two important changes to timeline recommendations: - New AI/SI ranking model that predicts probable time spent much better - Retrieval model trained to surface posts people read all the way through User active seconds is growing exponentially and hitting all-time highs almost

2K 2K 19K 3.0M 2K

Your agent's next upgrade might be its setup. I'd test that before retraining. Google Research's open-source Harness^2 refines prompts, tool guidance and skills while keeping model weights fixed. It tests edits on the current task and saves lessons to guide future refinement. How I'd try it: 1. Start with the local demo on Linux, Python 3.12 and uv. No model credentials needed; its editors are deterministic. 2. Try a small model-backed benchmark run. These runs incur API costs. 3. Inspect the saved harnesses, reflection notes and deliverables. The part I want: useful lessons that survive the task. #AIAgents #BuildInPublic github.com/google-researc…

@latent_node Yep. A kill switch that only lives in the UI is just a curtain. It has to sit on the handler.

Our real mistake: a live site's pay-per-use checkout was charging real money, and we had no way to pause just that path. Turning off the frontend also killed a seasonal purchase and the free trial. Claude Code tip: before you take payments, have the AI check these 7 things: 1. Right account: read-only call to /v1/account, match the id to your site map. Mismatch = stop. 2. Test mode on an isolated preview with its own test DB. 3. Prices come from the server, never the client. 4. Webhook signature verified, event IDs logged so duplicates are ignored. 5. Credits granted only from the webhook, never the success redirect. 6. Test cards: success, decline, 3DS, cancel, duplicate and delayed webhook. 7. One kill switch per paid entry point (pay-per-use, trial, subscription, one-off), plus a master switch that fails closed. At go-live, note your rollback deploy and watch the first real order. No credits in ~30 min? Flip that one switch. #ClaudeCode #Stripe docs.stripe.com/webhooks

@Rifat_EE An agent approving its own PR is just grading its own homework.

Stop reviewing every AI-written PR the same way. You'll burn out, and the risky ones still slip through. Anthropic's Claude Code team (Claude writes about 90% of that code) scales human review by risk. Copy it: Low: docs, copy, internal-only, additive behind a flag. 0 human approvals once CI and security checks are green. Medium: normal product changes, plus caching, auto-updater, system prompts. 1 person. High: the riskiest code. 2 people. Rules that make it safe: 1. The change sets the tier, not the author. 2. Auth, billing, CI and release workflows always need a person (CODEOWNERS + branch protection). 3. An agent's approval never counts as a human's. 4. Approval covers one commit. New commits, new approval. 5. Whoever prompted the agent owns the change. 6. Track reverts and incidents. If they rise, you didn't save time, you moved it. Their result: merged PRs more than doubled June to September, revert share down 40%+. #ClaudeCode #AIAgents claude.dev/blog/solving-c…

AI just got good at COBOL. The scariest bug it has to beat is one decimal place. OpenAI's new numbers: on its COBOL migration test, GPT-5.6 Sol got 5.4% of programs right. GPT-6.1 Sol gets 86.6% at low effort and 97.9% at higher effort. On MultiLangBench (build a SQLite-compatible engine in 20 languages), GPT-6 Astra averages 87.4%, with COBOL still its hardest at 72.1%. The part I'd tattoo on every payment codebase is their example of what fails: COBOL drops extra decimals unless the code says ROUNDED. So a fee of 1.449 is stored as 1.44. A "helpful" Python rewrite rounds it to 1.45. It compiles, it runs, the obvious tests pass, and it moves money wrong, one cent at a time. That's why OpenAI only counts a migration if every hidden test matches the old output byte for byte. That bar is right, and it's not just for banks. My 5 rules whenever AI touches money code: 1. Capture real outputs from the old code first, then diff the new code against them. 2. Store money as integer cents. No floats anywhere near a price. 3. Write the rounding rule into the prompt: truncate, half-up, or banker's. 4. Test the ugly edges: .005, refunds, zero-decimal currencies like JPY. 5. Any diff in a number, even 1 cent, gets a human look before it ships. Caveat: these are OpenAI's own evals. MultiLangBench is on GitHub, the COBOL migration set isn't public yet. x.com/OpenAI/status/…

OpenAI @OpenAI

3 days ago

Our latest model advances bring stronger capabilities for building and migrating software in low-resource programming languages such as COBOL, Fortran, and Visual Basic. These languages power software that banks, governments, and businesses depend on. Progress here means

826 685 10K 1.7M 2K

Your AI team can build the wrong thing beautifully. Saying no is still your job. #IndieDev

AI is most dangerous in the field you know least. Gergely Orosz nailed it today: when he uses AI on code he knows well, he spots the model BS-ing him instantly. Outside his lane, he trusts it. His example: he asked 3 AI models about a legal case he'd also paid a lawyer for. All 3 confidently said X. The lawyer said Y. The lawyer was right. That's Gell-Mann Amnesia with autocomplete. You catch the errors in your own field, then believe the next answer because it sounds just as sure. Vlad Mihalcea said the same from the coding side: Opus 5.5 mostly writes what he would, until it makes a design call that would fail in production. You only see that if you know production. My rules now: 1. In my lane: let AI go fast, I review the diff. 2. Outside it (legal, tax, payments, health): AI writes my questions, not my answers. 3. Ask for sources, then actually open them. 4. Have a second model attack the first answer before I act. 5. Money or legal risk still gets a human expert. AI just makes that call cheaper and better prepared. The model sounds the same either way. Your expertise is the only error detector. x.com/GergelyOrosz/s…

Gergely Orosz @GergelyOrosz

3 days ago

This messes with my mind ALL THE TIME When I do stuff with AI I am an *expert* in (eg coding on my project I know well, or doing research in an area I am very familiar with) I spot problems/errors/the model BS'ing me immediately. No big deal, I have it correct or start again

194 123 2K 139K 571

@melocalex The timestamp-at-the-top bug is so common it should have its own error message. Stable first, volatile last. Saving this.

Your agent is probably paying full price to reread the same instructions on every call. Check the cache before you switch models. Claude Opus 5.5 API, per million tokens: Input: $4 5-minute cache write: $5 Cache hit: $0.20 A hit costs 5% of normal input. Save this setup checklist: 1. Put reusable instructions first, changing task data after. 2. Keep that prefix identical and at least 512 tokens. 3. Add "cache_control": {"type": "ephemeral"} to its last content block. 4. Reuse it within 5 minutes. Each hit refreshes the timer. 5. Check usage.cache_read_input_tokens to confirm hits. 6. Compare total cost, writes and output included. No hits, no savings. Pick your most repeated task and measure it today. #AI #IndieDev platform.claude.com/docs/en/build-…

@Greg_GL_87 Approved twice, shipped once, debugged forever. What do you use as the second opinion at your agency?

LGTM now stands for "Looks Good To Model." Engineer 1 asks the AI to build it. Engineer 2 asks the same AI to review it. Two humans, one model, zero second opinions. The model just agreed with itself. I run a one-person company where AI agents write almost all of the code, so I hit this wall early. These are the review rules every release goes through now: 1. Payments, login and webhooks get a read-only adversarial review from a different model than the one that wrote the code. Its only job is to find what breaks. 2. The reviewer only comments. The author fixes. Max 3 rounds. 3. A preflight script runs first. If it fails, nobody reviews anything. 4. Code review and QA start on the same live preview at the same time, not just on the diff. 5. Nothing ships without screenshot evidence and a rollback ID written down first. 6. Anything that touches money still needs my personal sign-off. Same-model review is a spell checker grading its own spelling. Different blind spots are the whole point. x.com/Yuchenj_UW/sta…

Yuchen Jin @Yuchenj_UW

4 days ago

90% of software engineering in 2026: Engineer 1 → AI: “Build this and open a PR” Engineer 1 → Engineer 2: “Can you review?” Engineer 2 → AI: “Review this PR” Engineer 2 → Engineer 1: “LGTM” The irony: both engineers use the exact same AI model.

152 44 1K 153K 216