Adapter reference

Quick reference:

Tool	File(s)	Location
Shared / generic	`AGENTS.md`	repo root
Claude Code	`CLAUDE.md`	repo root (+ sub-dirs)
GitHub Copilot	`copilot-instructions.md`	`.github/`
Copilot file-scoped	`*.instructions.md`	`.github/instructions/`
Cursor	`*.mdc` rules	`.cursor/rules/`
Gemini CLI	`GEMINI.md`	repo root
Cline	`.clinerules`	repo root
Windsurf	`.windsurfrules`	repo root
Cursor legacy	`.cursorrules`	repo root

Startup — Detect Mode

Unless a mode was passed as an argument, always run detection first.

Step 1 — Scan for existing configs

Scan the repo for all known config files (see table above). Collect every file found.

Step 2 — Determine mode

Situation	Mode
No configs found	`init`
Configs found + argument says `add <tool>`	`add`
Configs found + argument says `check` or `drift`	`check` or `drift`
Configs found + argument says `edit`	`edit`
Multiple configs found, no argument	surface sync check first, then ask what user wants to do
One config found, no argument	ask: update, add a tool, or check drift?

Step 3 — Multi-config sync prompt

When two or more config files are found, before doing anything else:

"I found configs for [list tools]. Let me quickly check if they agree on the key things."

Run a lightweight cross-config comparison (see check mode below). Report any divergences immediately. Ask if the user wants to fix them before continuing to their original request.

Modes

`init` — No configs exist

Explain briefly: "You have no AI assistant config files yet. The most portable starting point is AGENTS.md — it's read by Claude Code, GitHub Copilot, Windsurf, and others without any tool-specific setup."
Ask: which AI tools does the team use? (multi-select)
- GitHub Copilot (default)
- Claude Code
- Cursor
- Gemini CLI
- Cline
- Windsurf
- Other (free text)
If called from repo-init with pre-filled context: skip this question and use the passed tool list.
Ask the preference questions (skip any already provided by repo-init):
- Shell preference (default: no preference): bash, zsh, fish, PowerShell, No preference
- Commit style (default: no preference): Conventional Commits, Gitmoji, No preference, Other
- Confirmation behavior (default: ask before destructive actions): Ask before destructive, Auto-approve safe / ask for destructive, Minimal confirmations
- Extra conventions (free text, optional): e.g. "never modify test files without asking", "always use named exports"
Decide which files to create:
- Always create AGENTS.md as the shared base.
- For each tool beyond the shared base, create the tool-specific file (see adapter reference for format).
- Show the file list and confirm before writing anything.
Write files. Use sentinel comments to wrap generated sections (see Adapter reference → Sentinel Format).
Confirm what was created and suggest next steps: add more tools (add), or run check after future tooling changes.

`add` — Onboard a new tool

Ask which tool to add (if not passed as argument). Show only tools that don't already have a config file.
Read all existing config files to extract current preferences (tools, conventions, methodology, shell).
Generate the new tool's config file using the extracted preferences + the adapter format. Wrap generated sections in sentinels.
Show a diff preview and confirm before writing.

`check` — Cross-config sync check

Goal: surface meaningful divergences between config files, not formatting noise.

For each config file, extract the following dimensions using the Adapter reference extraction hints:
- Named tools / frameworks (linter, test runner, formatter, runtime, package manager)
- Shell preference
- Commit style
- Any explicit methodology (e.g. TDD, SDD, spec-driven)
- Restricted or disallowed actions
Cross-compare all files on each dimension. Flag only where two files explicitly disagree (one says X, another says Y). Omissions are not flags — a file that doesn't mention the linter is not wrong.

For each divergence, show:

Divergence: linter
  CLAUDE.md        → Ruff
  copilot-instructions.md → Flake8
→ Which is current? [Ruff / Flake8 / Both intentional (suppress)]

For each resolved divergence: update all affected files (only within sentinel blocks or by appending if no sentinel exists). Never touch user-written content outside sentinels.
If no divergences: confirm "All configs agree on the key dimensions."

`edit` — Update a preference across all configs

Ask what the user wants to change (free text, e.g. "switch linter to Biome", "add conventional commits rule").
Identify which files contain the relevant preference.
For each file: locate the relevant line(s) within sentinel blocks. If the preference is inside a sentinel block, update it. If it's in user-written content, highlight it and ask before touching it.
Show a preview of all changes across all files. Confirm before writing.

`drift` — Tooling vs config consistency check

Goal: detect when the actual stack has changed but AI configs haven't caught up.

Load the full list of drift signal files and extraction rules from:

Drift signals reference

Procedure:

Scan for stack signal files: package.json, pyproject.toml, go.mod, Cargo.toml, *.eslintrc*, biome.json, ruff.toml, .tool-versions, .nvmrc, Dockerfile, etc. (full list in drift signals reference).
Extract named tools from each signal file (runtime version, linter, formatter, test runner, package manager).
For each AI config, extract the same dimensions (reuse check mode extraction).
Cross-compare: flag where a config mentions a tool that no longer appears in any stack signal file, or where a signal file mentions a tool that no config mentions.
Use git to add recency context:
```
git log -1 --format="%ar" -- <signal-file>
```
Only flag signal files changed more recently than the config that should reflect them.

Report findings as actionable items:

Drift detected:
  copilot-instructions.md says: Jest
  package.json now has: Vitest (changed 3 days ago)
→ Update copilot-instructions.md? [Yes / No / Suppress]

For each accepted fix: apply via edit mode logic. For suppressions: add an inline comment  immediately after the flagged line.
Structure drift — second pass (always run after tooling checks):

a. Stale path check — extract all path-like strings from each AI config (backtick-quoted or inline paths containing /). For each path, check whether it still exists in the repo. Flag missing paths at Medium severity.
```
Stale path in AGENTS.md:
  `packages/legacy-api/` — directory no longer exists
→ Remove or update this reference? [Update / Remove / Suppress]
```
b. Undocumented directory check — scan top-level and second-level dirs (for monorepos: packages/*, apps/*, services/*). Skip tooling/hidden dirs (node_modules, .git, dist, build, etc.). For each significant directory added more recently than the oldest AI config, check if any config mentions it. Surface undocumented ones at Low severity.
```
Undocumented directory: packages/payments-service/
  Added 5 days ago. No AI config mentions it.
→ Add a brief description to AGENTS.md? [Yes / No / Suppress]
```
If the user says Yes: ask for a one-sentence description, then insert it into the target config under a ## Project Structure sentinel block.

Full extraction rules, path patterns, skip lists, and the structure section template are in the Drift signals reference → Structure Drift.

Merge Safety Rules

These rules apply to every write operation across all modes:

Sentinel blocks are the safe zone. Generated content always lives inside:
```

...content...

```
For .mdc files (Cursor), use YAML comments: # ai-config:generated:<section> / # ai-config:end.
Never modify content outside sentinels unless the user explicitly asks.
First-time writes (no existing sentinels): wrap the entire generated file in a sentinel block so future syncs know what is safe to touch.
Suppression markers  are user-written and must never be removed.
Before any write: show the user a clear before/after diff and get confirmation.

Handoff Interface (from `repo-init`)

When repo-init calls this skill, it may pass a context object with pre-filled answers:

tools: [copilot, claude]
shell: bash
commit_style: conventional commits
confirmation: ask-before-destructive
extra: ["never modify test files without asking"]
methodology: spec-driven
stack:
  runtime: Node.js
  linter: ESLint + Prettier
  test: Vitest

When any of these keys are present, skip the corresponding question and use the provided value. Announce which values were pre-filled: "Using choices from repo-init: tools (Copilot, Claude Code), shell (bash)." so the user can correct anything before proceeding.

Guidelines

Detect before acting. Never assume what files exist — always scan first.
Prefer AGENTS.md for new setups. It has the broadest native support and avoids vendor lock-in.
Surface multi-file divergence immediately. When multiple configs exist, checking sync is always the first thing to offer.
Omission is not drift. A config that doesn't mention a tool is not wrong. Only flag explicit contradictions or stale mentions.
Never overwrite user content. Sentinel blocks define the safe zone. Everything outside is off limits without explicit user permission.
Ask one question at a time. Even during the init interview, keep each prompt focused on a single decision.
Preview before write. Always show what will be written and get confirmation before touching any file.
Suppressions are permanent. Once a user suppresses a drift signal, never re-raise it in future runs.
Be concise in reports. Show tool name, file, and the conflict — not paragraphs. Keep drift reports scannable.

Startup — Detect Mode

Step 1 — Scan for existing configs

Step 2 — Determine mode

Step 3 — Multi-config sync prompt

Modes

init — No configs exist

add — Onboard a new tool

check — Cross-config sync check

edit — Update a preference across all configs

drift — Tooling vs config consistency check

ai-config

AI Config Skill

Purpose

Platform Adaptation

Supported Config Files