{ "video": { "id": "uU5Gv2h8-9g", "title": "Claude Fable, Claude Tag, and Anthropic's Culture — Cat Wu & Thariq Shihipar ft Simon Willison", "duration": 3090, "upload_date": null, "channel": "AI Engineer", "source": "AI Engineer", "event_format": "In person" }, "analysis": { "video_id": "uU5Gv2h8-9g", "title": "Claude Fable, Claude Tag, and Anthropic's Culture — Cat Wu & Thariq Shihipar ft Simon Willison", "one_liner": "Anthropic's Cat Wu and Thariq Shihipar tell Simon Willison that Claude Tag — a multiplayer, proactive Claude living in Slack — now lands 65% of their product PRs, that frontier models let them cut the Claude Code system prompt by 80% by deleting examples and hard 'never do X' rules, and that auto mode (used internally since January, hardened against commissioned red teams) is how they run agents safely.", "summary": "The Claude Code team argues the bottleneck has moved from execution to taste: with idea-to-shipped down from six-to-twelve months to about a week, product and business sense matter more than implementation, and rewrites are now good because 'a codebase is a spec.' They detail how work actually ships at Anthropic — Claude Tag in Slack lands 65% of product PRs, code owners manually review the core while Claude code review fully reviews outer layers after a six-month trust-building process, and every incident PR becomes an eval. On prompting, they report a counterintuitive finding: frontier models (Fable, Opus 4.8) got better when examples and hard constraints were removed, producing an 80% token cut with a separate, still-full system prompt kept for older models.", "key_points": [ "Claude Tag, launched a week earlier in Slack, is multiplayer by default and proactive rather than reactive — you can tell it to monitor every bug report in a channel, open a PR, and tag the engineer who last touched that code, for the channel's lifetime. It has team memory (currently a markdown file per channel, shared by every Claude in it) and lands 65% of Anthropic's product-engineering PRs; Claude Code remains the place for complex interactive work.", "The Claude Code system prompt was cut by 80% in tokens for frontier models only — Anthropic now ships a different system prompt per model, with older models keeping the full version. Two changes drove it: removing examples (the model was more creative than the examples given) and replacing hard 'do not do this' constraints with context, because hard rules conflict with later user instructions and confuse Claude.", "Cat Wu's prompting rule: an instruction that is 90% true is a bug, because you give the prompt to the model 100% of the time. 'Always verify front-end changes' was softened after they realised a one-string copy fix with 'just make a quick fix' doesn't need verification.", "Code review is being moved off humans deliberately: code owners still manually approve core changes (the system prompt has one), but outer-layer changes are fully reviewed by Claude code review. They got there over 6+ months by measuring which files code review caught 100% of issues in, and by adding every incident-causing PR to an eval set so the fix never regresses.", "Auto mode has been in internal use since January; almost everyone at Anthropic uses it. A Sonnet classifier judges each tool/bash call against the conversation and your stated intent (so 'push this to GitHub' allows git push and 'don't push' denies it), and it also arbitrates escapes from the network sandbox. Thousands of evals plus multiple commissioned external red teams; they claim every issue found was mitigated, with evals to be published, and that residual risk is 'far lower than the average human reviewer.'", "Tool count is trending down, not up: grep and glob search tools were removed in favour of native bash, cardinality is kept low so each tool has a distinct function, and file edit survives mainly because it lets them render the approve-this-diff UI — Cat says they could probably drop it for auto-mode users. Thariq calls tool design 'more of a biology than a physics.'", "Thariq one-shot-prompted Fable to edit his ACM Agentic conference talk from raw files: it transcribed the video, noticed an auto-update popup in the deck recording and substituted his HTML deck source instead, dynamically cropped and tracked him as he paced the stage, and added animations — using ffmpeg and Remotion.", "Remote control was the surprise hit feature Cat didn't personally need: people plug in the laptop, close or lock the screen, and drive local Claude Code sessions from their phone on the couch. Also shipped: trusted devices, and credential injection so an agent can use a Datadog credential without holding or being able to read it (proxy swaps the token in on the fly).", "What Claude still can't do, per the speakers: design and UX taste — it follows the spec but 'the paddings might be off', it leans on existing app best practices, and there is a recognisable 'Opus aesthetic'; Cat wants future models to be interaction-design thought partners. Thariq wants it to reach into the real world and orchestrate science experiments." ], "takeaways": [ "Audit your system prompt for instructions that are only ~90% true and soften or delete them — and consider removing examples and hard prohibitions entirely when running frontier models, while keeping the verbose prompt for older or smaller models.", "Build the eval set before you need it: it's what lets a new model be a drop-in replacement (they ship Fable only once it's strictly better than Opus 4.8 across the suite), and it's where every incident-causing PR should land. Add behavioural evals alongside capability ones for the annoyances users actually complain about.", "Earn your way out of human code review file-by-file with measurement rather than switching it off — find the paths where automated review already catches 100% of issues, then remove the human there.", "Don't build your own AI Slackbot; a feedback channel your bot reads is a prompt-injection surface, and give agents their own provisioned identity and credentials so their actions are auditable.", "Put Claude in public channels, not DMs — Claude Tag's answer quality depends on being able to search every public channel, and the team found that watching colleagues prompt in the open raises everyone's usage quality and reduces slop." ], "topics": [ "coding-agents", "prompt-engineering", "evals", "code-review", "agent-security", "prompt-injection", "team-collaboration", "agent-memory" ], "tools": [ "Claude Code", "Claude Tag", "Claude Fable", "Claude Opus 4.8", "Claude Opus 4", "Claude Sonnet 3.7", "Haiku", "Claude Science", "co-work", "Anthropic", "Slack", "Bun", "Rust", "Zig", "Gemini API", "Seedance", "ffmpeg", "Remotion", "Datadog", "Mountain Project", "Airbnb" ], "quotes": [ { "text": "a codebase is a spec and maybe it's the only copy of the spec that you have", "at": "05:34", "url": "https://www.youtube.com/watch?v=uU5Gv2h8-9g&t=334s" }, { "text": "our internal version of claw tag lands 65% of our product PRs right now", "at": "08:19", "url": "https://www.youtube.com/watch?v=uU5Gv2h8-9g&t=499s" }, { "text": "you can soften the prompt such that it is actually 100% accurate because you are giving this prompt to the model 100% of the time", "at": "24:48", "url": "https://www.youtube.com/watch?v=uU5Gv2h8-9g&t=1488s" }, { "text": "the co-founders say like we don't negotiate against ourselves", "at": "45:53", "url": "https://www.youtube.com/watch?v=uU5Gv2h8-9g&t=2753s" } ], "words": 11869 }, "summary_url": "/#uU5Gv2h8-9g", "transcript": { "html": "/transcripts/uU5Gv2h8-9g.html", "txt": "/transcripts/uU5Gv2h8-9g.txt", "vtt": "/transcripts/uU5Gv2h8-9g.vtt" } }