Repo Intent Documenter

Build a grounded repository intent document that makes the project's purpose, assumptions, and unresolved questions obvious to future agents.

Decision Tree

What is the user asking for?

• Create or refresh a repo intent doc Run scripts/repo_intent_inventory.py, inspect the codebase, then write or update REPO_INTENT.md.

• Explain intent inline without writing files Use the same evidence model, but answer in chat and offer a doc path only if useful.

• Continue from an existing intent doc Read the existing doc first, preserve confirmed statements, and only revise claims contradicted by current evidence.

• User wants the doc opened after drafting Write the doc first, then open it with the available local mechanism. If no desktop opener exists, report the absolute path.

• The repository is missing or inaccessible Ask for the repo path or files. Do not invent intent from a project name alone.

Quick Reference

Need	Do
Fast repo map	python3 scripts/repo_intent_inventory.py <repo> --json
Evidence rules	Read references/methodology.md
Output shape	Start from templates/repo-intent.md and read references/document-contract.md
Ambiguity handling	Ask targeted questions after the draft, not before the first inspection
Failure modes	Read references/gotchas.md

Default Workflow

1. Resolve the repository root. Prefer the current working directory when it contains .git, a manifest, or existing project docs.
2. Run python3 scripts/repo_intent_inventory.py <repo> --json to collect the first pass of docs, manifests, tests, CI, entrypoints, and agent instructions.
3. Read the highest-signal files yourself: root README, AGENTS or CLAUDE files, package manifests, CI workflows, top-level source entrypoints, tests, examples, and existing docs.
4. Draft REPO_INTENT.md using templates/repo-intent.md.
5. Label claims as Certain, Strong inference, Tentative, or Open question. Every important claim needs an evidence anchor.
6. Ask the user only the questions needed to turn tentative claims into confirmed intent.
7. When the user answers, update the document and record the confirmation in the review log.
8. If requested, open the document for review after writing it.

Evidence Standard

Separate direct evidence from interpretation:

• Certain means a source file, README, manifest, test, or config states it directly.
• Strong inference means multiple independent signals point to the same intent.
• Tentative means the claim is plausible but based on weak or single-source evidence.
• Open question means the agent should not present the claim as true until the user answers.

Use file anchors whenever possible, such as README.md, package.json, src/server.ts, or tests/auth.test.ts. Include line numbers when the harness can provide them cheaply.

Output Path Policy

Default to REPO_INTENT.md at the repository root because future agents are most likely to discover a root-level intent file. Use docs/repo-intent.md only when the repository clearly keeps all durable project docs under docs/ and root-level docs would violate local conventions.

Do not automatically edit AGENTS.md, CLAUDE.md, or other persistent agent instructions unless the user asks. Instead, add a short "Suggested integration" note in the intent doc when linking it from agent instructions would help.

Question Protocol

Ask questions after the first draft exists, so the user can correct a concrete artifact.

Keep questions:

• evidence-linked: mention the file or signal that created the ambiguity
• answerable: prefer a short choice or factual confirmation
• prioritized: ask the smallest set that would materially improve future agent behavior
• non-blocking: leave unresolved questions in the doc if the user is not ready to answer

Gotchas

1. Do not turn directory names into product strategy. src/app proves structure, not customer intent.
2. Do not flatten uncertainty. A confident but unsupported sentence is worse than an explicit open question.
3. Do not ask broad discovery questions before reading the repo. The point is to make the agent do the first pass.
4. Do not overwrite confirmed human intent just because current code is incomplete.
5. Do not bury the actual purpose under a long architecture inventory. Architecture supports the intent; it is not the intent.

Reading Guide

Need	Read
Full inspection method, signal ranking, and confidence labels	references/methodology.md
Required document sections, destination rules, and review loop	references/document-contract.md
Pitfalls and recovery patterns	references/gotchas.md
Starting Markdown shell	templates/repo-intent.md