Generating AGENTS.md templates

AGENTS.md is a dedicated, predictable place to store agent-focused project context (project structure, commands, testing recipes, and conventions) that would clutter a human-oriented README. Prefer concrete, repo-derived facts over generic advice.

Inputs you MUST use

You have access to all files in the repository. Use them.

Prioritize these sources (if present):

• AGENTS.md (existing), README.md, CONTRIBUTING.md, docs/
• Build/task runners: Makefile, package.json scripts, justfile, tox.ini, noxfile.py
• Tooling configs: pyproject.toml, setup.cfg, .pre-commit-config.yaml, .editorconfig
• CI: .github/workflows/*, .gitlab-ci.yml, azure-pipelines.yml, etc.
• Release docs: CHANGELOG.md, release scripts, tagging docs

Output contract

Produce one Markdown document: the full contents of AGENTS.md.

• If an AGENTS.md already exists: preserve useful content, update stale parts, and keep the same intent/tone.
• If you cannot find evidence for a required detail, insert a single-line marker: > TODO: <what’s missing and where to confirm it>
• Do not invent commands, versions, paths, or policies.

Authoring workflow (do this in order)

1. Inventory & evidence

   • Identify the project’s language(s), key frameworks, and tooling by reading configs.
   • Locate canonical commands (lint/format/test/type-check/build) from task files.
   • Identify where tests live and how CI runs them.

2. Draft the outline

   • Use the exact section order in “AGENTS.md template” below.
   • Keep it “medium”: enough to execute common work correctly, but not a full handbook.

3. Fill from repo facts

   • Prefer verbatim commands (copy exactly).
   • Reference files by relative paths (e.g., pyproject.toml, mtrust/calls.py).

4. Plan-first compatibility (no interactive approvals)

   • Do not write “ask first / require confirmation” directives.
   • Instead, encode risk as plan-time requirements, e.g.:
     • “When proposing a plan that changes public interfaces, call out breaking-change risk.”
     • “When changing dependencies, include rationale and impact in the plan.”

5. Quality checks

   • Ensure required sections are present (repo map, testing, conventions, dependency/tooling).
   • Ensure Markdown only (no HTML), and keep lines ≤ 100 chars where practical.
   • Ensure TODOs are minimal and actionable (point to where to look).

AGENTS.md template (keep exact order)

1) # AGENTS.md

• One paragraph: what the repo is, what it does, and who this file is for (agents + devs).

2) ## Repository Structure

• Show a top-level tree or bullet map of directories and key files.
• For each item: one sentence purpose.
• Only drill into subdirectories when they are critical entry points.

3) ## Build & Development Commands

• Group commands under short headings (e.g., Lint/Format/Test/Type-check/Build).
• Use fenced bash blocks.
• Copy commands exactly as found (Makefile targets, scripts, etc.).
• If multiple equivalent ways exist, list the canonical one first and mention alternatives.

4) ## Dependency & Tooling Notes

Include only what you can prove from the repo:

• Language/runtime versions supported (e.g., python = ">=3.11").
• Framework versions/ranges when pinned or constrained.
• Key tooling (linters/formatters/test runners/type checkers) and where configured.

5) ## Code Style & Conventions

• Formatting/linting rules at a high level (line length, import order, formatter).
• Naming conventions (modules, tests, settings).
• Used changelog conventions (if present).

6) ## Testing Recipes

• Where tests live (paths), and how to run:
  • Full suite
  • A single file / a single test / keyword filtering (use the repo’s framework)
  • Coverage (if present)
• Note any rules for external calls:
  • If tests mock network/third-party APIs, state where fixtures/mocks live.
  • If unclear, add a TODO.

7) ## Maintenance Notes

• A short list of what must be kept current as the repo changes:
  • Update this file when structure/commands/tooling change.
  • Update CHANGELOG.md per the present changelog conventions.

Writing rules

• Be specific: prefer “pytest + pytest-django” over “run tests”.
• Avoid duplication with README; link/point to it when appropriate.
• No marketing language; focus on operational accuracy.
• Do not mention any tool-specific behavior (no references to Codex CLI, etc.).