Archive (@ArchiveExplorer)

Anthropic published exactly how they get Claude to work autonomously for 6 hours: • 0:00 - the two failure modes killing every long agent run • 0:09 - the feature_list.json format + "unacceptable to remove tests" rule • 0:16 - incremental progress + why they use Puppeteer MCP for testing • 0:22 - the fixed session-start protocol every fresh agent runs • 0:28 - verbatim: a typical session transcript • 0:32 - the failure modes → solutions cheat sheet read it today, then the article below shows how to steal every piece for your own harness.

David Ha, CEO sakana ai & ex-head of google brain tokyo. recorded a 58-min english segment inside a japanese TBS interview. AND IT'S THE BEST AI TALK OF THE LAST 2 MONTHS. taxonomy: - MCTS across a portfolio - fire the same prompt at 5 models, race them, verifier picks the winner and continues from there - ARC-AGI-2 - the reasoning benchmark that broke every frontier model. sakana ai hit new sota on it with the MCTS approach - the agent writes code, runs it, reads the error, dispatches more code harness lesson: one big model brute-forcing = fragile. many small models + a verifier = leverage. stack: gpt-5 + gemini 2.5 + DeepSeek-R1 + Qwen 3 → MCTS race → new sota on ARC-AGI-2. buried on a japanese business channel while everyone argues about context windows. nobody clipped it. watch it today, then read the article below ↓

met an anthropic engineer making $1.2M a year. asked him how he ships alone at the pace of a full team. he didn't answer. sent me his .claude/. one folder. SAME MODEL - DIFFERENT RESULT. everyone's still picking between opus and sonnet like the model is the ceiling. it isn't. the real lever is what the model wakes up into: CLAUDE.md → hooks → verifier subagent → skills → mcp → memory → shift notes. you stop chatting with the model. you write the folder once. the folder runs the model. - CLAUDE.md - the contract - settings.json - the permissions - hooks/ - the reflexes - agents/verifier - the shift-notes cop - skills/ - 33 muscle memories - .mcp.json - the tools - MEMORY.md - the shift log that's the stack. full breakdown in the article below. bookmark before he realizes i posted it.

A guy runs AI ads for local roofers and they pay him $2,000-$5,000 a month each he doesn't film anything. He doesn't hire anyone he opens Facebook's Ad Library and finds ads that have been running for 6+ months - those are the proven winners then he drops them into Higgsfield and has it rebuild the same ad with the client's logo claude writes the script and the ad copy for him in one afternoon he's got 10-15 video ads for a roofing company that never ran a single ad before he launches them on Meta and leads start coming in through a simple form an AI bot texts every lead back in under 2 minutes and books them straight onto the client's calendar the roofer just wakes up to booked appointments. He has no idea AI did all of it some of these clients have paid him the same fee every month for 2 years straight he bills per lead or a flat retainer - either way it's recurring, and the AI does most of the work one skill, a few tools, and small local businesses that have never touched advertising save this, the quietest AI money right now isn't apps - it's running ads nobody knows are AI