agents-md

AGENTS.md generator

Generates the workspace-root AGENTS.md and CLAUDE.md redirect shim from a .code-workspace scan, with optional Compose-backed runtime blocks in existing project instructions.

Inputs

• The target root must contain a *.code-workspace file (multi- or single-folder). If none exists, stop and do not generate files — the Project Matrix is built from its folders list; tell the user to rerun from a folder holding one. If several exist, ask which one to use before scanning.
• Never read or copy the agent's own global or user instruction files (~/.claude/CLAUDE.md, ~/.codex/, global AGENTS.md, personal memory/rules). Generate rules from the shipped templates and workspace evidence. Zero attribution: omit co-author, AI, and tool attribution from generated files and all output.

Rules

• Workspace-root AGENTS.md is the single source of truth for Codex CLI, Claude CLI, Antigravity CLI, Cursor CLI, Opencode CLI, and GitHub Copilot CLI; shared operating instructions — context, memory policy, issue routing, skill use — live there.
• CLAUDE.md is only a redirect shim for Claude CLI: emit assets/claude-md-template.md byte-for-byte, nothing more, and never read other context from it.
• Create exactly one root pair. Project synchronization edits existing AGENTS.md files only; never create project instruction files.

Modes

Default: generate or refresh the root pair, and synchronize existing project runtime blocks when .devcontainer/compose.yml exists. Read Project runtime synchronization for matching, migration, and approved edits.

A request naming only project-runtime synchronization selects runtime-only: follow that reference and existing stack detection; skip root generation, migrations, and setup suggestions. Never create or edit workspace-root AGENTS.md or CLAUDE.md in this mode. Root-generation rules and completion checks below apply only in the default mode.

Workspace scan

Scan every workspace folder first: note any VISION.md/vision.md at the workspace or a project root (feeds North star) and detect each stack. Multiple stacks → the intro slot says so and warns against mixing conventions or code across projects.

Stack detection

The Stack cell is fact read from the app-root manifest and lockfiles — never guessed from a folder's name, a root-level manifest, or a file's mere existence. Find the real app root first — often nested (application/, app/, src/) — and read the manifest there. Per-ecosystem manifest reads (PHP, frontend build, JS/TS, Python, test runners, non-app folders): references/stack-detection.md.

Project Matrix format

One row per .code-workspace folder, in folders order — never invent or drop projects. Columns exactly | Project | Path | Stack |:

• Project: the PROJECT-CODE from the folder name — strip emojis, uppercase, collapse punctuation/space runs to single hyphens, trim leading/trailing hyphens (Payments API → PAYMENTS-API); path basename when name is missing; type prefix for non-app folders (PACKAGE-QUEUE, DB). Name the full PROJECT-CODE from the Project Matrix everywhere; never mix one project's conventions, tokens, or components into another.
• Path: the folder path relative to the workspace root; . marks the meta/workspace row.
• Stack: one terse line — language, every primary framework, build tooling, package manager; versions only when a manifest pins them; non-app folders state the role. No prose in cells.

If normalization yields an empty or duplicate PROJECT-CODE, stop and ask for a unique code before emitting the matrix. Inspect a . folder's manifest too; it can be an application root.

Sample:

| Project      | Path             | Stack |
| :---         | :---             | :---  |
| PAYMENTS-API | ../payments-api  | PHP 8.3 / Laravel 13 / Vite / Composer + npm (application/) |
| DB           | ../db-migrations | MySQL / raw SQL migrations |

npm comes from application/package-lock.json; Vite alone never implies a package manager.

Emitted skeleton

Emit assets/agents-md-template.md byte-for-byte, filling only the bracketed slots (intro line, Project Matrix, skills tables + startup note, ### Runtime tool-calling tables, and the #### North star list when emitted), except customizations retained under Versioning and regeneration. Three rules bind the default skeleton:

• Emit ### Matt skill routing only when Matt Pocock's skills are installed: a SKILL.md for ask-matt, grill-with-docs, or to-spec exists under .agents/skills/ or .claude/skills/ at the workspace root or in the user's home. Otherwise delete the subsection. If this check and the runtime's skill list disagree, ask. The gradient table lists external skills either way.
• Emit #### North star only when the scan found a vision file, listing each with its scope; never fabricate one or restate its content — the subsection points at the file. None found → delete it.
• Keep Issue titles exactly as concise as the skeleton has it — routing/label procedure lives in the issue skills, never here — and leave the two Context & native memory placeholders unfilled.

Working with skills

Generate the gradient and companion tables from references/skills-manifest.md — the single source; never hardcode rows here. Kit entries with phase: companion render only in the companion table.

Apply the manifest's note exclusions before emitting any catalog or startup note.

Gradient cells list each skill as `/name` followed by its manifest note, comma-separated, in manifest order.

Fill the [RUNTIME TOOL-CALLING …] slot from references/tool-calling.md, following its Emitting into AGENTS.md format exactly; open a per-runtime table (claude, codex, copilot, cursor, opencode, antigravity) only when a cell is missing or unclear. Native memory setup, on request only: memory defaults. Never link the reference files (they do not ship) or restate the Local orchestration rule.

Versioning and regeneration

The skill version is v47. Both generated root files carry the marker <!-- agents-md marker · v47 · re-run /agents-md to regenerate --> as their first line (the first line of each template asset in assets/). Bump it here and in both template assets together whenever these rules change. Legacy pre-v6 markers: regeneration.

On run, check for an existing workspace-root AGENTS.md:

• None — generate fresh.
• Marker present — before overwriting, show the complete diff section by section — never truncated or summarized — with customizations flagged separately, and confirm. If the user does not respond, stop without writing and say so.
• No marker (hand-authored) — do not rewrite it; show what generation would add or change, merge only user-approved sections, and let the existing file win everywhere else.
• Keep PROJECT-CODEs stable — keep each existing row's code, matched by Path. A code that would now derive differently (a renamed folder) is a separate diff item needing explicit approval; never change it silently. New folders get derived codes.
• Preserve user edits — carry over user-filled placeholder values (especially the GLOSSARY.md and specs/adr paths), customized rule bodies, and foreign sections, per regeneration.
• Write safely — re-read each target just before writing; if it changed since approval, show the recomputed diff again. On any write or move failure, stop and report what was and was not written.
• Regenerate the CLAUDE.md shim only if it is missing or its marker is stale.

Migrations (ask first)

List every move and every file whose links would be rewritten — scanning only the artifacts root — and ask first. Declined or no response → keep the existing paths; never rename unattended.

• An artifact subfolder still under docs/ (pre-v7): docs to specs.
• CONTEXT.md or CONTEXT-MAP.md still in use (pre-v47): glossary rename.

Output

Chat carries the pre-write diff, the migration move list and report, the manifest file read for each Project Matrix row, the rule-link check result, competing instruction files, and the close — suggesting env-guard hooks when hosts are named — suggest /setup-matt-pocock-skills, and /design-system for each UI project, then stop — suggest only, never run them.

Completion criteria

• Requested project synchronization satisfies the matching, approval, and preservation checks in its reference
• AGENTS.md sections appear in template order; ### Matt skill routing present only when Matt's skills resolve, #### North star only when a vision file was found
• Every in-document link target exists as a heading or <a id> in the output — the check is reported in chat; omitting optional sections preserves all numbered rules and Runtime tool-calling
• Regeneration: foreign sections, user-filled values, and rule-body customizations survive; any customized-rule replacement has explicit approval — or the run stopped without writing
• Generated catalogs and startup notes contain no manifest-excluded entries; kit companions appear only in the companion table
• Worktree routing survives Matt routing omission; when emitted, Matt routing passes the verified checkout to the task skill
• Project Matrix row count equals the folders count; chat names the manifest read for every Stack cell; existing PROJECT-CODEs unchanged unless a change was approved
• Both generated files open with the current version marker; Context & native memory placeholders intact or carried over filled
• diff of CLAUDE.md against the shim template is empty
• After an approved migration, moved artifacts have no stale docs/<sub>/ links, only listed files were rewritten, and unrelated docs/ content and declined moves remain intact
• A search of both files finds no session data: dates, conversation references, machine-local absolute paths