Agent MD Wizard

Problem

In monorepos, mixed-language repositories, or projects with sparse documentation, agents often do not know which build, lint, or test commands to run, and they may modify generated files, execute risky migrations, or bypass team review rules. Asking a model to guess repository conventions usually produces guidance that misses command boundaries and can accidentally overwrite manual adjustments in an existing AGENT.MD.

How It Works

This skill first runs scripts/detect_repo_context.py to inspect manifests, lockfiles, Docker or CI files, and existing AGENT.MD signals, then returns normalized JSON. It follows the six-round question tree from references/question-tree.md: project shape and topology, main stack and runtime, install/dev/build/test commands, coding and testing expectations, risk controls, and collaboration notes. High-confidence detections are used to prefill answers, while low-confidence fields are asked explicitly. The final draft is previewed through scripts/build_agent_md.py; the file is written only after confirmation into the repository root as AGENT.MD.

Boundaries

It is intended for generating or incrementally updating an agent policy for one repository, not for replacing CI, code review, or security scanning. If command signals are weak, key commands need manual confirmation. For mixed-language repositories, short branch questions can be triggered to avoid merging submodule commands into the root. When an AGENT.MD already exists, prefer incremental edits over rewriting the document tone.