Claude Code instructions

@AGENTS.md

Treat AGENTS.md as the canonical repository contract. Follow it for both production work and code changes; do not duplicate or weaken its rules here.

Starting a task

• For a request to create, recut, capture, render, or judge a demo, read SKILL.md completely before editing files or running provider-backed stages.
• Read only the task-relevant production references named in AGENTS.md.
• Inspect existing product evidence before proposing a story. Ask a focused question only when the audience, intended change, or required evidence cannot be discovered safely.
• Keep new working projects under projects/<slug>/ so footage, narration, renders, and QA outputs remain local and ignored.
• Prefer the ./pdd commands in AGENTS.md. Use individual tools/*.py stages when diagnosing or intentionally controlling one stage of the documented pipeline.
• Do not invoke paid providers, capture authenticated pages, or render media merely to explore the repository. Run those stages only when the user's requested production requires them.

Claude-specific behavior

• This root file loads automatically in Claude Code. Use /memory if you need to verify which instruction files are active.
• ./pdd install-skill --target claude is optional and only needed to expose Product Demo Director while Claude works from another repository. Do not install it without the user's request.
• Never paste provider keys into chat or shell history. Read local configuration through the documented environment variables without printing values; verify a required key exists before a provider-backed stage.
• Do not declare a film complete from a visual impression or model score. Report the exact final media path and the deterministic preflight/QA result required by AGENTS.md.

For a typical launch request, interpret “make a demo” as: inspect evidence, identify one meaningful change, author a source-bound single-focus story, capture missing states, synthesize or stage narration, pace cuts between thoughts, render, finish, run strict QA, and then optionally judge.