CLAUDE.md — Claude Code Entry Point¶

This file is the Claude Code-specific entry point for this repository. It does not restate the full operating contract — read AGENTS.md first. Everything in AGENTS.md applies to Claude Code sessions.

Start here¶

Read AGENTS.md in full before making any change.
Load the applicable venture's ventures/<slug>/AGENTS.md when the task is scoped to a venture (e.g., Shelfery, Digital Products).
Follow the load order in docs/ai/context-loading-order.md — do not skip layers or load them out of order.

Planning¶

For any nontrivial task, write a work plan in work/active/ before editing files, per docs/ai/planning-and-work-logs.md. Trivial single-file fixes do not require a formal plan but still need a brief explanation of what changed and why.

Editing discipline¶

Make targeted edits only. Do not perform unrelated rewrites, reformatting, or cleanup outside the scope of the current task.
If you notice unrelated problems while working, note them (work log or docs/governance/open-questions.md) rather than fixing them inline.

Validation before completion¶

Run python scripts/check.py (and python -m unittest discover tests where relevant) before considering any task complete. Report the result. Do not claim validation passed if you did not actually run it.

Contradictions and assumptions¶

If instructions in this repository conflict with each other, or with the task you've been given, surface the contradiction explicitly rather than silently resolving it in one direction. State any assumption you had to make in plain language, in the work log or in the response to the user.

ADRs for architecture decisions¶

Any material, hard-to-reverse architecture or policy decision made in the course of a task must be recorded as an ADR under docs/decisions/ using the template in templates/, as part of the same task.

Source-backed claims¶

Any factual claim you add to a document (market data, competitor behavior, technical facts about a third-party platform) must be attributable to a real, checkable source, or explicitly labeled as an assumption/hypothesis. Do not present model-generated guesses as researched fact.

Hard prohibitions¶

Never commit secrets, API keys, customer credentials, or private customer/personal data to this repository.
Never fabricate test results, benchmark numbers, or validation output. If you did not run something, say so.
Never change a document's status field to approved. Only a human can approve policy; agents may draft and propose (status: draft or status: proposed) but never approve.
Never begin product implementation work ahead of required specifications unless a human has explicitly authorized skipping the gate for that specific task.

When in doubt about scope, authority, or whether something is a "material" decision, stop and ask rather than guessing.

Report an issue about this page