Tailor

Tailor

The framework's environment-triggered customization engine. Once an adopter's stack is detectable — manifests and lockfiles committed — /tailor proposes the configuration the framework should take: filled tech-strategy.md golden paths, REVIEW.md/CLAUDE.md review steering, and a prune list of unused framework pieces.

/tailor proposes only. It never silently writes to .claude/rules/, .claude/settings.json, CLAUDE.md, or any other tracked config file. Every run ends in a reviewable plan at scratchpad/tailor-proposal.md — nothing lands in a tracked file until the user approves it (see Output Contract). This promise is now hook-backed at the ask tier: pre-tool-use-validator.sh asks for confirmation before any direct Write/Edit to these same paths, rather than leaving the propose-only contract as convention alone.

Detect

Scan for stack signals per references/detection.md, the mechanical signal-to-golden-path-row contract. Cover:

• Language manifests + lockfiles: package.json/pnpm-lock.yaml/yarn.lock, pyproject.toml/uv.lock, go.mod, Cargo.toml, Package.swift/*.xcodeproj, build.gradle(.kts)
• Tool configs: biome.json, .eslintrc*, [tool.ruff] in pyproject.toml, .golangci.yml
• CI: .github/workflows/*
• Infra: Dockerfile, *.tf, Railway/Fly configs

Produce a stack-fingerprint table, one row per claim, with an evidence file for each:

Precedence when signals conflict, highest wins: committed lockfile > manifest field > dependency presence > inference. Never assert a claim without citing the file it came from — no vibes-based stack claims.

Assess

Read .claude/rules/tech-strategy.md. Classify every golden-path section as one of:

• placeholder — matches the shipped example verbatim (the file's own "Customization Required" banner marks this state)
• already-customized — differs from the shipped example, no detected conflict
• conflicting-with-detection — differs from the shipped example AND contradicts what Detect found

Never overwrite an already-customized row without flagging it first. A conflicting row becomes a flagged decision in the proposal, not a silent replacement.

Propose: Fill

For each detected language, draft a ready-to-paste golden-path table using the project's actual choices, with versions pulled from manifests where present (e.g. "react": "^19.2.0" in package.json → React 19.2). For each undetected language section, propose deleting it — a prune candidate, not a silent removal.

Propose: Steer

Invoke the review-steering skill's workflow to draft REVIEW.md and verify CLAUDE.md's review-context section is current. Include both the drafted REVIEW.md content and any CLAUDE.md delta in the proposal — do not write either file directly.

Propose: Instantiate

For each pack in .claude/templates/stack-packs/ whose stack Detect actually found, adapt the exemplar to the fingerprint — substitute package manager, test runner, framework, and version strings for the detected reality, evidence-cited per substitution; flag a default when no signal exists rather than inventing one. See references/packs.md for the discovery and adaptation rules this phase follows.

Render golden-path.skill.md at its adopter-repo target path (e.g. .claude/skills/typescript-golden-path/SKILL.md) and ci-gates.yml as a ready-to-paste block for the adopter's own CI workflow — never for a stack Detect did not find.

Propose: Prune

List framework pieces irrelevant to the detected stack: unused golden-path language sections, hook checks that can't fire for this stack, skills the adopter may want to gate. Each entry needs a one-line rationale and the exact file/line — no bulk "remove unused stuff" claims.

Output Contract

/tailor proposes only — application is a separate, user-approved step, never automatic. A single proposal file at scratchpad/tailor-proposal.md contains:

1. Stack-fingerprint table (Detect)
2. Per-file proposed blocks, ready to paste (Fill, Steer, Prune)
3. Rendered pack files (Instantiate), each with its adopter-repo target path and the evidence-cited substitutions applied
4. Conflict flags (Assess)
5. An apply checklist — one checkbox per file to change

Application happens only after the user approves the plan — either by manually pasting the approved blocks, or by re-invoking /tailor with the approved checklist.

Modes

• all (default) — Detect + Assess, then all four Propose phases; full proposal
• detect — Detect only; fingerprint table
• fill — Detect + Assess + Propose: Fill
• steer — Detect + Assess + Propose: Steer
• prune — Detect + Assess + Propose: Prune
• instantiate — Detect + Assess + Propose: Instantiate

Constraints

• NO silent writes to .claude/rules/, .claude/settings.json, CLAUDE.md, or any other tracked config file — proposal only, always
• NO deleting or overwriting user-customized content — flag the conflict instead
• NO stack claims without cited evidence — every fingerprint row names its source file
• NO rendering a pack for a stack Detect did not find — instantiation is scoped strictly to detected stacks, never speculative
• ALWAYS re-run after major stack changes (new language, package-manager migration, framework swap) — a stale proposal is worse than none
• ALWAYS require explicit user approval before anything in the proposal is applied — the constraint most tempting to skip under time pressure

Handoffs

• To review-steering: Propose: Steer delegates REVIEW.md/CLAUDE.md drafting to its workflow
• To /swarm-plan: when detection reveals structural work beyond a config fill (e.g. a full framework migration)
• From first install: docs/customization.md's "Configure Your Tech Stack" step points here once the repo has a detectable stack

$ARGUMENTS