init — Claude Code skill: write and prune AGENTS.md

---
name: init
description: Creates, updates, or optimizes an AGENTS.md file for a repository with minimal, high-signal instructions covering non-discoverable coding conventions, tooling quirks, workflow preferences, and project-specific rules that agents cannot infer from reading the codebase. Use when setting up agent instructions or Claude configuration for a new repository, when an existing AGENTS.md is too long, generic, or stale, when agents repeatedly make avoidable mistakes, or when repository workflows have changed and the agent configuration needs pruning. Applies a discoverability filter—omitting anything Claude can learn from README, code, config, or directory structure—and a quality gate to verify each line remains accurate and operationally significant.
metadata:
  tags: initialization, agents, context-engineering, agents-md, maintenance
---

## When to use

Use this skill when creating or updating `AGENTS.md` for a repository.

Use it especially when:
- the current `AGENTS.md` is long, generic, or stale
- agents repeatedly make the same avoidable mistakes
- repository workflows changed and agent guidance needs pruning

## Instructions

Treat `AGENTS.md` as a **living list of non-discoverable landmines and workflow gotchas**, not a codebase overview.

### Core rule: discoverability filter

Before adding any line, ask:

> Can an agent discover this by reading the repo (`README`, code, config, scripts, directory tree)?

- If **yes**: do **not** include it in `AGENTS.md`.
- If **no**, and it materially affects task success/cost/safety: include it.

### What earns a line

Include only guidance that is:
1. **Non-discoverable** from repository files alone
2. **Operationally significant** (changes commands, outcomes, or safety)
3. **Actionable** (specific enough to execute)

Typical examples:
- Non-standard tooling choices (e.g. use `uv` instead of `pip`)
- Command caveats (e.g. tests must run with `--no-cache` due to fixture behavior)
- Hidden constraints/landmines (deprecated directories still imported in production)
- Critical local conventions that are not encoded in lint/tests/config

### What to remove or avoid

Do **not** include:
- Tech stack summaries
- Directory structure overviews
- Architecture descriptions agents can infer from code
- Generic best-practice advice
- Rules already enforced by tooling (linters, typecheck, tests, CI)
- Mandatory boilerplate headers unless the repo explicitly requires one

### Recommended structure

Prefer short, high-signal sections such as:
- `Scope & routing` (which areas need separate/module-local AGENTS files)
- `Non-discoverable commands`
- `Landmines / do-not-touch areas`
- `Task-specific constraints`

For large repos, recommend **hierarchical AGENTS.md** files near relevant modules instead of one monolithic root file.

### Source files to check first

- Existing `AGENTS.md`
- `README.md`
- `PROJECT.md` (if present)
- Cursor rules (`.cursor/rules/` or `.cursorrules`)
- Copilot instructions (`.github/copilot-instructions.md`)
- `GEMINI.md`
- CI/workflow files and package manager config (for command/tooling mismatches)

If `AGENTS.md` exists, improve it incrementally instead of replacing it blindly.

### Maintenance mindset

`AGENTS.md` is temporary guidance, not permanent configuration.

When recurring issues appear:
1. Prefer fixing the root cause in code/tooling (lint rule, test, script, structure)
2. Keep only the minimum instruction needed until the root cause is solved
3. Prune stale instructions aggressively

### Quality gate before finalizing

For each line in `AGENTS.md`, verify:
- Is it non-discoverable?
- Is it still accurate today?
- Does it materially reduce mistakes/cost/time?

Delete any line that fails one of these checks.