Anthropic's AI-native SDLC playbook, run on a real repo: the six files exactly as the run produced them, the three hook scripts, the trap eval, and the order to add them tonight
The companion to the video. On September 6, 2026 we ran all six stages of Anthropic's AI-native SDLC playbook (claude.com/blog, Aug 21 2026), headless, on typicode/json-server (75,704 GitHub stars) against its real open issue #1618, a --delay flag open since November 2024. The templates in this PDF are the exact files that run produced, not examples. PAGE 1 — what this is, the credit line, and the order to adopt tonight: intent.md, then CLAUDE.md, then the protect-paths hook, then one trap eval, with one line each on what it stops; then spec.md, plan.md, REVIEW.md and bands.yaml when a team appears. PAGE 2 — intent.md: the template headings, the filled example condensed from the run, and the two paste-ready prompts (the interview, then 'now write intent.md'). PAGE 3 — spec.md and plan.md section skeletons, the playbook's spec prompt verbatim, and the plan-mode one-liner (claude -p ... --permission-mode plan > plan.md). PAGE 4 — CLAUDE.md verbatim (Commands with the one verify command and its healthy output, Conventions, Architecture, Things Claude gets wrong) plus .claude/settings.json verbatim and the shell-write gap that release-gate.sh closes. PAGE 5 — the three hook scripts verbatim: protect-paths.sh, release-gate.sh, format.sh. PAGE 6 — the trap eval JSON verbatim, what run.sh and check.sh check, agent-evals.yml verbatim, and REVIEW.md verbatim with the -d -1 crash the review pass found. PAGE 7 — bands.yaml verbatim, how detect.sh escalates (1 sigma log, 2 sigma diagnose read-only, 3 sigma propose), the diagnoser prompt, and the bot-written intent.md that called the drill a drill. PAGE 8 — the honest box: the bare prompt also shipped working code on this small task in 94 seconds for $0.97; the chain (about 10 minutes, about $11) bought scope, guardrails and an audit trail, not correctness; n = 1. Independent run from Hyperautomation Labs — not affiliated with Anthropic.
Free. No spam. Unsubscribe anytime.