Agentify project

Converge a project — greenfield or half-way-there — toward the canonical agent-neutral structure specified in references/target-structure.md (read it before the audit), so a team of developers using different agents shares one lean, progressively disclosed setup. Re-running converges further and reports drift.

Invocation: /agentify-project [audit] [path]

• audit — run step 1 only, print the findings and what a full run would change, modify nothing; the periodic drift check.
• path — target a subdirectory instead of the repo root (see Monorepos).

Principles

• Ask, don't enforce: prefer asking — all options shown, one recommended — over enforcing a choice, except for obvious easy wins (e.g. clear progressive-disclosure moves). Questions answerable from the codebase go to the codebase.
• One approval gate per destructive/bulk change; show diffs or proposals before applying; nothing happens behind the user's back. Never commit — suggest reviewing the diff and leave committing to the user.
• Generic: behavior is driven by references/agent-matrix.md — read it whenever agent detection or an agent's specifics are needed. One documented exception, the Claude carve-out: a team with no Claude users at all is rare, so Claude wiring (the .claude/skills/ links, the Claude gitignore entries) is applied unconditionally, never detection-gated. A CLAUDE.md pointer is not part of it: current Claude Code reads AGENTS.md itself (target-structure §2).

1. Audit (read-only)

• Detect each target-structure area's state, and which agents are in use by their traces (matrix).
• Warn — never block — if the git working tree is dirty, so the user can commit first and keep the restructure reviewable as an isolated diff.
• Detect nested entry files (monorepo packages).
• Report any CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md in a parent directory of the main checkout, except the user-level ~/.claude/CLAUDE.md: with no CLAUDE.md pointer in the repo it makes Claude Code skip AGENTS.md (target-structure §2), so the pointer becomes a menu item. The main checkout is the parent of git rev-parse --path-format=absolute --git-common-dir, not --show-toplevel, which in a linked worktree (agents often create one inside the repo, e.g. <repo>/.claude/worktrees/<name>/) returns the worktree. The main checkout's own root CLAUDE.md, .claude/CLAUDE.md and CLAUDE.local.md are the repo's files, never a parent-directory CLAUDE.md that stays, and never alone keep or recreate a pointer.
• Diff duplicated entry files (agents without include syntax) against AGENTS.md; divergence becomes a re-sync menu item — the one artifact class guaranteed to drift.
• Check every path and link in the entry files and index resolves; a dead one is a menu item (fix or drop).
• Detect Claude-only assets (.claude/commands/, .claude/agents/, .claude/rules/): report, leave. Report unscoped rules files as always-on, and flag .claude/rules/ content as candidate governing-doc material (path-scoped rules are a Claude-specific disclosure mechanism); legacy commands may get a one-off note that they could become skills — the conversion itself is out of scope.
• Record always-on line counts — entry files plus any rules file without a path scope (they load every session): menu annotations now, coarse before-figure for /context-checkup later.

Done when every target-structure area has a detected state and every agent trace found is accounted for.

2. Menu

Multi-select over the applicable change areas only — entry-file indirection, AGENTS.md slimming into docs, skills relocation + symlinks, gitignore hygiene, planning dir, onboarding doc, README pointer — each annotated with audit findings (e.g. "CLAUDE.md, 180 lines, no AGENTS.md → indirection + slimming"). "All" is the first and recommended option. Report already-converged areas ✓ and omit them from the choices. Every area is independently skippable: a team may deliberately stay Claude-only (no multi-agent support) — then apply the selected areas within the Claude-only variant in target-structure.md.

3. Execute — live, step by step

Apply each accepted area immediately, behind its approval gate — no migration-plan artifact: the moves are mostly mechanical, and re-running converges. Suggested order: entry files → slimming/docs → skills → gitignore → extras.

Entry files. Invariant covering every starting state: the content of all existing entry files — any agent's, including a divergent CLAUDE.md + AGENTS.md pair — feeds the slimming mapping; AGENTS.md ends canonical; every detected third-party entry file ends as a pointer (or duplicated content, per matrix); CLAUDE.md ends absent or as a pointer, per target-structure §2. Common case (CLAUDE.md exists, no AGENTS.md): rename it to AGENTS.md with a plain filesystem move (no git mv; git detects the rename at commit time) and recreate CLAUDE.md as @AGENTS.md only when §2 calls for one (a genuinely Claude-specific remainder, sessions that cannot load AGENTS.md, or a parent-directory CLAUDE.md that stays); otherwise it stays deleted. Run from a linked worktree: a tracked entry file the branch renames or replaces gets a report note that worktree sessions keep loading the main checkout's copy until the branch reaches it; an untracked one (the gitignored CLAUDE.local.md) gets target-structure §2's CLAUDE.local.md remedy, with no such note. AGENTS.md-only and third-party-only repos: same invariant — pointers only for the third-party agents detected, never a bare CLAUDE.md; seed AGENTS.md from whatever entry-file content exists. Greenfield (no entry file of any agent): minimal seed from cheap reconnaissance (README, package manifest / build files) — one-liner, verified build/test/run commands, planning-dir convention; deep doc generation is out of scope, the structure grows later via /self-improve and /memory-doctor.

Slimming (mapping-first). Draft one full mapping: each block → stays / moves to existing doc X / new doc Y, applying the keep-vs-extract test (target-structure §1); where a block's relevance is verifiable from the codebase, check the code instead of asking. Present the mapping with obvious calls pre-decided; walk only doubtful blocks as individual questions, each with a recommendation, worded via explain-in-simple-language when available; one approval gate before any edit. Offer a compaction pass on blocks that stay; write extracted docs compact from birth — via compact-docs-writer when available, else inline with its core principle (least text, zero information loss). Each extracted doc gets a plain-link index entry in AGENTS.md with its "read when…" hook.

Skills. Move real dirs, create relative symlinks (committed), per matrix; leave foreign/personal links untouched and report them.

Gitignore. Show the proposed block as a diff; apply on approval.

Extras. Planning dir + AGENTS.md convention line; onboarding doc, indexed with its hook (target-structure §1); README pointer.

Done when every selected area is applied or explicitly skipped by the user.

4. Verify (mechanical)

Every AGENTS.md reference resolves; symlinks point at real dirs; git check-ignore confirms each ignore entry; no CLAUDE.md, .claude/CLAUDE.md or CLAUDE.local.md shadows AGENTS.md (each absent, or an @AGENTS.md pointer — a tracked file the branch replaces in the main checkout is not reported; a real file above the main checkout or an untracked shadowing file still is) and, when the root keeps a CLAUDE.md, every nested AGENTS.md has its pointer too (target-structure §2); duplicated entry files match AGENTS.md. Close with a short converged-state report of what changed; after a slimming, state the hook test: give a fresh session a task that needs one extracted doc and check it opens it; a miss means a weak "read when…" hook, fixed via /self-improve when available (gated as in §5). Point at the onboarding doc's copy (target-structure §8) only when that doc exists at the end of the run.

5. Close

Suggest, each only when available (look in the agent's skills and plugin dirs, as manual-only skills are missing from the in-session list) and applicable, with this skill's own namespace prefix if any — skip silently otherwise:

• /memory-doctor — when the agent accumulates project memory (e.g. auto memory) — to relocate saved lessons into the new docs structure and share them with the team; it no-ops harmlessly on empty memory, so no content check is needed.
• /context-checkup — measure what now auto-loads, with the audit's always-on line counts as the coarse before-figure.

Monorepos

Full treatment targets the repo root. Nested entry files are offered only the cheap mechanical indirection (nested CLAUDE.md → nested AGENTS.md; pointer per target-structure §2 — always kept while the root has a CLAUDE.md); defer nested slimming to a re-run with that subdirectory as the path argument. A path run covers that directory's entry-file and slimming areas only; skills, gitignore, planning dir, onboarding doc and README pointer stay root-level.