Imported from ATTCKDigital/smith (skills/smith/SKILL.md). Install upstream with npx skills add ATTCKDigital/smith --skill smith. Copyright stays with the author.

SpecKit Project Initialization

Bootstrap a project with the full SpecKit spec-driven development workflow. This skill scans the existing codebase to pre-fill answers, interviews you about your project, and generates all configuration files.

Arguments: $ARGUMENTS

Workflow

Phase 0: Locate Skill Assets

The SpecKit skill assets (templates, scripts, commands, agents) are bundled at:

~/.claude/skills/smith/
├── templates/          # spec, plan, tasks, checklist, agent-file templates
├── scripts/            # Shell scripts for feature management
├── commands/           # Slash command definitions (smith.specify, etc.)
└── agents/             # Agent definitions (architect, senior-qa, etc.)

Verify this directory exists. If missing, abort with: "SpecKit skill assets not found at ~/.claude/skills/smith/. Please reinstall the skill."

Phase 1: Preset Check

If $ARGUMENTS contains --preset <name>, load preset defaults and skip to Phase 3 (confirmation). Available presets:

If no preset, proceed to Phase 2.

Phase 2: Codebase Detection

Scan the current working directory to detect existing project characteristics. For each signal found, record it as a pre-filled answer. Present findings as a detection report before asking questions.

2.1 Project Identity Detection

• Check for existing CLAUDE.md — extract project name and description if present
• Check for package.json → extract name, description
• Check for pyproject.toml → extract [tool.poetry].name, description
• Check for Cargo.toml → extract [package].name, description
• Check for go.mod → extract module name
• Check for README.md → extract first heading and description paragraph
• Check git remote URL for project name hints

2.2 Tech Stack Detection

Frontend:

• package.json dependencies: react → React, vue → Vue, svelte → Svelte, next → Next.js, @angular/core → Angular
• package.json dependencies: tailwindcss → Tailwind, styled-components, @emotion/react, sass
• tsconfig.json exists → TypeScript; otherwise JavaScript
• Check for Storybook: .storybook/ dir or @storybook/* in devDeps

Backend:

• pyproject.toml or requirements.txt: flask → Flask, fastapi → FastAPI, django → Django
• package.json dependencies: express → Express, fastify → Fastify, hono → Hono
• go.mod → Go
• Cargo.toml → Rust
• Gemfile: rails → Rails, sinatra → Sinatra

Database:

• pyproject.toml/requirements.txt: sqlalchemy → SQLAlchemy, psycopg2/asyncpg → PostgreSQL, pymongo → MongoDB
• package.json: prisma → Prisma, typeorm → TypeORM, mongoose → MongoDB, pg → PostgreSQL, better-sqlite3/sqlite3 → SQLite
• Look for *.db or *.sqlite files → SQLite
• Look for docker-compose.yml services: postgres, mysql, redis, mongo

Package Managers:

• poetry.lock → Poetry; Pipfile.lock → Pipenv; requirements.txt → pip
• package-lock.json → npm; yarn.lock → Yarn; pnpm-lock.yaml → pnpm
• Cargo.lock → Cargo; go.sum → Go modules

Auth:

• Search for oauth, jwt, passport, flask-login, next-auth, auth0 in dependencies or code

2.3 Tooling Detection

• Linting: .eslintrc* or eslint.config.* → ESLint; ruff.toml or [tool.ruff] in pyproject.toml → Ruff; .pylintrc → Pylint
• Formatting: .prettierrc* → Prettier; Ruff format → Ruff
• Testing: vitest or jest in package.json → Vitest/Jest; pytest in pyproject.toml → Pytest; go test patterns → Go test
• CI/CD: .github/workflows/ → GitHub Actions; .gitlab-ci.yml → GitLab CI; Jenkinsfile → Jenkins
• Docker: Dockerfile or docker-compose.yml → Docker
• Storybook: .storybook/ directory → Storybook

2.4 Structure Detection

• frontend/ + backend/ → Web app (separate dirs)
• src/ only → Single project
• apps/ or packages/ → Monorepo (directory heuristic)
• ios/ or android/ → Mobile
• Check for existing .specify/, .claude/commands/, .claude/agents/

2.4.1 Monorepo Tool Detection

Detect monorepo orchestration tools by their config files. These take precedence over directory heuristics:

Config file detection:

• pnpm-workspace.yaml → pnpm workspaces
• package.json with workspaces field → npm/yarn workspaces (parse JSON to confirm)
• nx.json → Nx monorepo
• turbo.json → Turborepo
• lerna.json → Lerna
• rush.json → Rush

When monorepo config is found:

1. Set structure type to "Monorepo" with high confidence
2. Record the orchestration tool (e.g., "pnpm workspaces", "Nx", "Turborepo")
3. Parse workspace configuration to identify package locations:
   • pnpm: Parse packages array from pnpm-workspace.yaml
   • npm/yarn: Parse workspaces field from package.json (can be array or object with packages key)
   • Nx: Parse projects from nx.json or scan for project.json files; check workspaceLayout for custom paths
   • Turbo: Note presence; Turbo typically layers on top of npm/yarn/pnpm workspaces
   • Lerna: Parse packages array from lerna.json; check useWorkspaces flag
   • Rush: Parse projects array from rush.json
4. Identify the monorepo root (where the config file lives) — this may differ from cwd if running from a sub-package
5. Store detected workspace globs for use in CLAUDE.md generation

Multiple tools: If multiple configs exist (e.g., turbo.json + pnpm-workspace.yaml), record both — Turbo is often used alongside a workspace manager.

2.5 Existing Configuration Detection

• Check for existing CLAUDE.md — warn user it will be regenerated (offer to preserve or replace)
• Check for existing .specify/ — warn and offer to update or replace
• Check for existing .claude/commands/ — same
• Check for existing .claude/agents/ — same
• Check for existing constitution at .specify/memory/constitution.md

2.6 Present Detection Report

Display findings in a clear table:

## Codebase Detection Report

| Category | Detected | Confidence | Value |
|----------|----------|------------|-------|
| Project Name | Yes | High | "my-project" (from package.json) |
| Structure | Yes | High | Monorepo |
| Monorepo Tool | Yes | High | pnpm workspaces (from pnpm-workspace.yaml) |
| Workspace Packages | Yes | High | packages/*, apps/* |
| Frontend | Yes | High | React 19 + TypeScript |
| Styling | Yes | High | Tailwind CSS |
| Backend | Yes | High | Flask 3.x + Python 3.12 |
| Database | Yes | Medium | SQLite (found .db files) |
| Package Manager | Yes | High | pnpm (from pnpm-lock.yaml) |
| Auth | No | — | Not detected |
| Linting | Yes | High | ESLint + Ruff |
| Testing | Yes | High | Vitest + Pytest |
| CI/CD | Yes | High | GitHub Actions |
| Docker | Yes | High | Docker Compose |
| Storybook | No | — | Not detected |
| Default Branch | Yes | High | development (from origin/HEAD) |

Items marked "No" or "Medium" confidence will be asked about in the interview.

Default-branch detection step: Detect the repository's default/integration branch and record it for the report row above and the base-branch interview question. Run:

git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's|refs/remotes/origin/||'

If this resolves (e.g. development, develop, trunk), use it as the detected value; if it fails (no origin remote, detached HEAD), fall back to main. This value seeds the base-branch interview question (below) and is written to the constitution as base_branch:.

Phase 3: Generate Init Intake Document

Instead of prompting the user interactively, generate a persistent intake document at specs/init-intake.md.

IMPORTANT: This file is the permanent record of project configuration decisions. Never delete it. If it already exists, read it and use the existing answers — do not regenerate unless the user explicitly asks to refresh it.

3.1 Create specs/ directory if it doesn't exist

mkdir -p specs

3.2 If specs/init-intake.md does NOT exist, generate it

Create specs/init-intake.md with:

• A frontmatter block with generation date, project name, and status
• The full Codebase Detection Report table from Phase 2
• All 29 questions organized in 5 groups (Project Identity, Tech Stack, Development Tooling, Quality Standards, Workflow Preferences)
• For each question: the options, a Recommended answer with rationale based on codebase detection, and an Answer field pre-filled with the recommendation

Base-branch question (Workflow Preferences group): Include a question — distinct from the feature-branch naming question (Q25) — that asks which branch is the project's integration/base branch. Smith cuts feature branches from it and targets PRs at it. Wording: "Which branch is your integration/base branch? (Smith cuts feature branches from it and targets PRs at it.)" Auto-detect the recommended value and pre-fill the Answer with it:

git symbolic-ref refs/remotes/origin/HEAD 2>/dev/null | sed 's|refs/remotes/origin/||'

If detection fails (no origin remote, detached HEAD), recommend main. The user may override the recommendation with any branch name (e.g. a development → main promotion flow where the user targets development even though origin/HEAD points at main). The chosen value is written to the constitution frontmatter as base_branch: in Phase 4.

• A Project-Specific Notes section capturing any important context detected during the scan that doesn't fit the standard questions (e.g., multi-service architecture, N8N workflows, custom patterns)

The recommended answers should be based on:

1. High-confidence codebase detections (use detected values directly)
2. Best practices for the detected stack (when no detection available)
3. Project context from memory, STATUS.md, or other project docs

3.3 If specs/init-intake.md ALREADY exists, read and use it

Parse the existing file to extract the Answer lines for each question. These answers drive all subsequent file generation in Phase 4. Tell the user: "Found existing specs/init-intake.md — using your recorded answers. Edit the file and re-run /smith to change any decisions."

3.4 Interactive Walkthrough

After generating (or loading) the intake document, walk the user through every question interactively, one at a time. For each question, present:

1. Question number and text (e.g., "Q5. Frontend framework")
2. Context — why this decision matters for the project
3. Options — each option with a brief explanation of trade-offs
4. Recommended answer — your recommendation with reasoning based on codebase detection and project context
5. Current answer — what is currently recorded in the intake file

Then wait for the user's response. Accepted responses:

• A specific choice → update the Answer line in specs/init-intake.md immediately
• "ok" / "yes" / pressing enter with no input / confirming the recommendation → keep the current answer, move to next question
• A question or comment → answer it, then re-present the same question
• "skip" → keep current answer, move to next question
• "back" → return to the previous question

After all questions are answered, present a final summary table of all answers and ask the user to confirm before proceeding to Phase 4.

Update specs/init-intake.md in real-time as the user answers — do not batch updates.

Phase 4: Generate Files

After the interactive walkthrough is complete and the user confirms, generate all project files.

4.1 Scaffold Project Directories

Create the directory structure and copy files from the skill assets:

mkdir -p .specify/memory .specify/templates .specify/scripts/bash
mkdir -p docs/sessions
mkdir -p specs/questions
mkdir -p .smith/vault/sessions .smith/vault/agents .smith/vault/queue .smith/vault/bank .smith/vault/ledger
mkdir -p .smith/vault/active-workflows

# Bootstrap marker so the workflow-gate hook (PreToolUse) allows the
# rest of /smith init to write the remaining scaffold. Without this,
# the gate would deny every subsequent Write/Edit because no marker
# exists yet. Cleared at the end of /smith init (see Phase 4.10).
cat > .smith/vault/active-workflows/bootstrap.yaml << EOF
workflow: smith
feature: project-initialization
# NOTE: this captures the CURRENT HEAD branch (the branch /smith init runs on),
# not the project's integration/base branch. The `|| echo main` is the
# current-branch fallback for a detached HEAD — it is intentionally NOT the
# configured base branch and must not be replaced with get-base-branch.sh.
branch: $(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo main)
started: $(date -u +"%Y-%m-%dT%H:%M:%S")
EOF

The docs/sessions/ directory holds session chat logs (timestamped Q&A records with YAML frontmatter for searchability). The specs/questions/ directory holds structured question files generated before complex changes (numbered questions with options, recommendations, and answer fields). Both are required by the global workflow tenets in ~/.claude/CLAUDE.md.

The .smith/vault/ledger/ directory holds the Ledger — Smith's learned knowledge from past workflow executions. After creating the directory, scaffold the Ledger template files: patterns.md, antipatterns.md, tool-preferences.md, edge-cases.md, project-quirks.md (each with a header and empty-state message), and meta.yaml (initialized with creation date and zero counters). See the smith-reflect skill for the exact file formats.

Copy from ~/.claude/skills/smith/:

• templates/* → .specify/templates/
• scripts/* → .specify/scripts/bash/

This scripts/* glob copies every helper in skills/smith/scripts/, including get-base-branch.sh and clear-active-workflow.sh — no per-script copy instruction is needed, and get-base-branch.sh is scaffolded to .specify/scripts/bash/get-base-branch.sh automatically.

Make scripts executable:

chmod +x .specify/scripts/bash/*.sh

4.2 Generate Constitution (.specify/memory/constitution.md)

Generate a constitution file using the interview answers. Structure:

---
base_branch: [from the base-branch interview answer; default `main` if unanswered]
---

# [Project Name] Constitution

## Core Principles

### I. Code Quality Standards
[Generated based on detected linters, languages, component patterns]

### II. Testing Standards
[Generated based on detected test frameworks, coverage target from Q20]

### III. User Experience Consistency
[Generated only if frontend detected — based on styling framework, component tool]

### IV. Performance Requirements
[Generated from Q21 answers, or standard defaults]

## Technology Constraints
[Generated from tech stack answers Q5-Q13]

## Development Workflow

### Branch Strategy
[Generated from Q25 (feature-branch naming) and the base-branch question. State the
configured integration/base branch — the value written to `base_branch:` in the
frontmatter above — so it is human-readable here as well as machine-readable in the
frontmatter that `get-base-branch.sh` parses.]

### SpecKit Workflow
[Standard — always included]

### Commit Standards
[Generated from Q26, Q28]

### Quality Gates
[Generated from Q23, Q27]

## Governance
[Standard boilerplate with project name, version 1.0.0, current date]

Use the following rules for content generation:

• Python backend: Include Ruff linting, Pydantic validation, snake_case naming
• JavaScript/TypeScript frontend: Include ESLint, PascalCase components, camelCase functions
• React + Storybook: Include component directory pattern, Storybook-first development
• Tailwind: Include "Tailwind utility classes required; custom CSS MUST be justified"
• Docker: Include container-related standards
• SQLAlchemy: Include "No N+1 queries" rule
• Monorepo: Include workspace-aware rules (see below)
• Only include sections relevant to the detected/selected stack

Monorepo-specific constitution rules (only if monorepo detected):

• Include "Changes affecting multiple packages require cross-package testing"
• Include "Shared packages must maintain backwards compatibility or coordinate breaking changes"
• Include workspace dependency guidelines (internal vs external)
• For Nx: Include affected-based testing rules (nx affected:test)
• For Turborepo: Include pipeline caching expectations
• For pnpm: Include workspace protocol usage (workspace:*)

4.3 Generate CLAUDE.md

Generate a CLAUDE.md file at the project root using the interview answers. Structure:

# CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

## Project Overview
[Project name and description from Q1-Q2]

### Project Governance
See `.specify/memory/constitution.md` for binding project principles.

### Business Domain
[Domain from Q3, entities from Q4 if provided]

## Tech Stack

### Frontend
[From Q5-Q7, only if frontend exists]

### Backend
[From Q8-Q12, only if backend exists]

### Monorepo Structure
[Generated only if monorepo detected — include workspace layout, package locations, and orchestration tool]

## Git Commit Guidelines
[Generated from Q26, Q28 — include linking format, co-author trailer]

## Common Commands

### CI/CD Scripts
[Generated based on detected CI scripts, or suggest creating them]

### Docker Development
[Generated only if Docker detected]

### Frontend Development
[Generated from frontend stack — npm/yarn commands]

### Backend Development
[Generated from backend stack — poetry/pip commands, test commands]

### Database
[Generated from database choice — migration commands if ORM detected]

### Monorepo Commands
[Generated only if monorepo detected — include workspace-aware commands]

## Architecture

### Backend Structure
[Generated from directory scan or standard template for the stack]

### Frontend Structure
[Generated from directory scan or standard template for the stack]

### Key Patterns
[Generated based on stack — e.g., "Application Factory" for Flask, "API Routes" for Next.js]

## SpecKit Workflow
[Standard section — always included, customized with review gate selections from Q27]

### Issue-Driven Workflow
[Standard — customized with issue tracker from Q24]

### Subagent Usage
[Standard — always included]

### Review Gates (Mandatory)
[Generated from Q27 selections]

### When Asked "What's Next?"
[Standard — always included]

### Directory Structure
[Standard .specify/ structure]

### Branch Naming
[From Q25]

## E2E Testing with Playwright MCP
[Generated only if Playwright MCP selected in Q19]

IMPORTANT rules for CLAUDE.md generation:

• Only include sections relevant to the stack (no Docker section if no Docker, no Storybook section if no Storybook, etc.)
• Generate actual commands based on the real package manager and frameworks
• If existing CLAUDE.md was found and user chose to preserve, merge new sections with existing content
• Include the SpecKit Workflow section verbatim — this is the standard workflow
• Do NOT add a ## Recent Changes / changelog section, and do NOT @-reference a doc that any workflow then appends per-change prose to (e.g. a data-model.md with a growing "Last Updated" header). @-referenced files are loaded into context in full, every session, and per-change appends turn them into multi-hundred-KB context sinks that also cause merge conflicts on every branch. The durable per-change record already lives in .smith/vault/sessions/ + the Ledger + git — those ARE the changelog. Keep CLAUDE.md and any @-referenced doc small and stable: structural/durable content only (overview, tech stack, ERD, entity tables, enums), never accumulating history.

Monorepo-specific CLAUDE.md content (only if monorepo detected):

For the ### Monorepo Structure section, include:

• Orchestration tool name and version (e.g., "pnpm workspaces", "Nx 17.x", "Turborepo")
• Workspace root location (usually cwd, but note if different)
• Package locations parsed from config (e.g., packages/*, apps/*)
• Brief description of each discovered package/app if parseable

For the ### Monorepo Commands section, generate tool-specific commands:

pnpm workspaces:

pnpm install                    # Install all workspace dependencies
pnpm -F <package> add <dep>     # Add dependency to specific package
pnpm -F <package> run <script>  # Run script in specific package
pnpm -r run build               # Run build in all packages
pnpm -F "...<package>" test     # Test package and its dependents

npm/yarn workspaces:

npm install                     # Install all workspace dependencies
npm -w <package> run <script>   # Run script in specific package
npm run <script> --workspaces   # Run script in all workspaces

Nx:

npx nx run <project>:<target>   # Run target in specific project
npx nx affected:build           # Build only affected projects
npx nx affected:test            # Test only affected projects
npx nx graph                    # Visualize project dependencies
npx nx run-many -t build        # Run build in all projects

Turborepo:

turbo run build                 # Run build with caching
turbo run test                  # Run tests with caching
turbo run build --filter=<pkg>  # Build specific package
turbo run build --filter=...<pkg>  # Build package and dependents

Lerna:

lerna bootstrap                 # Link local packages
lerna run build                 # Run build in all packages
lerna run test --scope=<pkg>    # Run test in specific package
lerna publish                   # Publish changed packages

Rush:

rush install                    # Install dependencies
rush build                      # Build all projects
rush build -t <project>         # Build specific project and deps
rush test                       # Run tests

4.4 Set Up .claude/commands/

Create the project's .claude/commands/ directory and copy command files from the skill assets:

mkdir -p .claude/commands

Copy from ~/.claude/skills/smith/commands/:

• All smith.*.md files → .claude/commands/
• review-respond.md → .claude/commands/

4.5 Set Up .claude/agents/ (Optional)

If the user selected review gates in Q27, create custom agent definitions:

mkdir -p .claude/agents

Copy from ~/.claude/skills/smith/agents/:

• If architect gate enabled: architect.md → .claude/agents/
• If product-manager gate enabled: product-manager.md → .claude/agents/
• Always copy: senior-qa.md, staff-backend.md, staff-frontend.md, staff-fullstack.md, staff-infrastructure.md

After copying, replace [PROJECT_NAME] placeholders in the agent files with the actual project name from Q1.

4.6 Create .claude/settings.json (if not exists)

If no .claude/settings.json exists, create one with sensible defaults:

{
  "permissions": {
    "allow": [
      "Bash(git status:*)",
      "Bash(git log:*)",
      "Bash(git diff:*)",
      "Bash(git branch:*)",
      "Bash(./.specify/scripts/*:*)",
      "Bash(.specify/scripts/bash/clear-active-workflow.sh:*)",
      "Bash(.specify/scripts/bash/get-base-branch.sh:*)"
    ]
  }
}

The clear-active-workflow.sh entry is listed explicitly so Smith workflow cleanup still works on projects that add Bash(rm:*) to the deny list as a safety rail. The helper is narrow (single file, no globs, path-escape guarded) and lets skills remove the active-workflow marker without weakening the deny rule.

The get-base-branch.sh entry is listed explicitly for the same reason: every Smith skill that needs the integration branch resolves it via .specify/scripts/bash/get-base-branch.sh, so projects with strict Bash allow/deny rails must permit it. The helper is read-only (no writes, no globs, no recursion) and always echoes a branch name (main on any non-resolution).

Add framework-specific permissions based on detected stack:

• npm projects: "Bash(npm run:*)", "Bash(npm install:*)"
• Poetry projects: "Bash(poetry run:*)", "Bash(poetry install:*)"
• Docker projects: "Bash(docker:*)", "Bash(docker compose:*)"
• CI scripts: "Bash(./scripts/*:*)"

Add monorepo-specific permissions based on detected orchestration tool:

• pnpm workspaces: "Bash(pnpm:*)", "Bash(pnpm -F:*)", "Bash(pnpm -r:*)"
• Nx: "Bash(npx nx:*)", "Bash(nx:*)"
• Turborepo: "Bash(turbo:*)", "Bash(turbo run:*)"
• Lerna: "Bash(lerna:*)", "Bash(npx lerna:*)"
• Rush: "Bash(rush:*)"

If .claude/settings.json already exists, do NOT overwrite — inform the user they may want to add the .specify/scripts permission manually, plus Bash(.specify/scripts/bash/clear-active-workflow.sh:*) if their config includes a broad Bash(rm:*) deny rule (needed so Smith workflow cleanup can remove active-workflow markers), and Bash(.specify/scripts/bash/get-base-branch.sh:*) so base-branch resolution works under strict Bash rails.

4.7 Create .gitignore Entries

If .gitignore exists, append SpecKit-related entries if not already present:

# SpecKit / Claude Code
.claude/agent-memory/
.claude/todos/

If .gitignore doesn't exist, create one with standard patterns for the detected stack plus the above entries.

Then merge the canonical Smith vault & index policy (the team-shareable selective policy from feature 36). This is a single source of truth shipped as a template asset — do NOT inline-duplicate the list here; read it from the installed path and merge idempotently using sentinel markers.

Source template (prefer the installed copy, fall back to the repo-dev copy):

• ~/.claude/skills/smith-index/templates/.gitignore-smith-additions
• (fallback) skills/smith-index/templates/.gitignore-smith-additions

And the matching .gitattributes policy (de-emphasizes generated .meta in diffs):

• ~/.claude/skills/smith-index/templates/.gitattributes-smith-additions
• (fallback) skills/smith-index/templates/.gitattributes-smith-additions

Idempotent sentinel merge contract:

• The template block is wrapped in # >>> smith-gitignore-policy >>> … # <<< smith-gitignore-policy <<< (and # >>> smith-gitattributes-policy >>> … # <<< smith-gitattributes-policy <<< for .gitattributes).
• If the project file already contains the opening sentinel, replace everything between the opening and closing sentinel (inclusive) with the fresh template block.
• Otherwise append the whole block (creating .gitattributes if it does not exist).
• Lines OUTSIDE the sentinels are never touched.

Staleness guard (Feature 38). A pre-Feature-36 installed template is the inverse policy (it IGNORES index/files/ + index/systems/) and lacks the # >>> smith-gitignore-policy >>> sentinel. /smith init has no upstream clone to auto-repair from (unlike /smith-update), so if the resolved installed template is stale it must NOT be applied. Prefer the repo-dev copy if it is current; otherwise warn the user to run /smith-update and SKIP the policy merge rather than writing the wrong (inverted) policy.

Run this self-contained python3 block (robust sentinel replace-or-append — no sed portability issues):

# Resolve the template source (installed path first, repo-dev fallback).
SENTINEL='# >>> smith-gitignore-policy >>>'
SMITH_TPL_DIR="$HOME/.claude/skills/smith-index/templates"
[ -f "$SMITH_TPL_DIR/.gitignore-smith-additions" ] || SMITH_TPL_DIR="skills/smith-index/templates"

# Staleness guard: if the resolved template lacks the sentinel, it's a pre-Feature-36
# (inverted) policy. Try the repo-dev copy; if that's also stale/absent, skip the merge.
if ! grep -qF "$SENTINEL" "$SMITH_TPL_DIR/.gitignore-smith-additions" 2>/dev/null; then
    if grep -qF "$SENTINEL" "skills/smith-index/templates/.gitignore-smith-additions" 2>/dev/null; then
        SMITH_TPL_DIR="skills/smith-index/templates"
    else
        echo "WARNING: the installed Smith gitignore template is stale (pre-Feature-36,"
        echo "         no sentinel marker) and no current repo-dev copy was found. Skipping"
        echo "         the .gitignore/.gitattributes policy merge to avoid writing the wrong"
        echo "         (inverted) policy. Run /smith-update to refresh the installed template,"
        echo "         then re-run /smith init (or just /smith-update, which re-merges the policy)."
        SMITH_TPL_DIR=""
    fi
fi

if [ -n "$SMITH_TPL_DIR" ]; then
python3 - "$SMITH_TPL_DIR" <<'PYEOF'
import sys, pathlib

tpl_dir = pathlib.Path(sys.argv[1])

def merge(target_path, tpl_path, open_marker, close_marker):
    tpl = pathlib.Path(tpl_path)
    if not tpl.exists():
        print(f"WARNING: template not found: {tpl_path} — skipping")
        return
    block = tpl.read_text().rstrip("\n") + "\n"
    target = pathlib.Path(target_path)
    existing = target.read_text() if target.exists() else ""

    if open_marker in existing and close_marker in existing:
        # Replace the managed region (inclusive) with the fresh block.
        pre = existing.split(open_marker, 1)[0]
        post = existing.split(close_marker, 1)[1]
        new = pre.rstrip("\n") + ("\n\n" if pre.strip() else "") + block + post.lstrip("\n")
        action = "refreshed"
    else:
        sep = "" if (not existing or existing.endswith("\n\n")) else ("\n" if existing.endswith("\n") else "\n\n")
        new = existing + sep + block
        action = "appended"
    target.write_text(new)
    print(f"{action}: {open_marker.split()[1]} in {target_path}")

merge(".gitignore", tpl_dir / ".gitignore-smith-additions",
      "# >>> smith-gitignore-policy >>>", "# <<< smith-gitignore-policy <<<")
merge(".gitattributes", tpl_dir / ".gitattributes-smith-additions",
      "# >>> smith-gitattributes-policy >>>", "# <<< smith-gitattributes-policy <<<")
PYEOF
fi   # end: SMITH_TPL_DIR non-empty (merge not skipped by staleness guard)

The policy is a negative list: .smith/ is NOT ignored wholesale. Only the specific volatile paths in the block are ignored; everything else under .smith/ (the manifest.md, config/, index/files/ + index/systems/ .meta describe layer, ledger/, bank/, agents/, and sessions/*.md) is committed and shared with the team. Re-running /smith init replaces the managed region in place, so the block never duplicates.

4.8 Optionally Scaffold System Specs

Smith v2 path-resolver tier 1 reads .specify/systems/<id>/spec.md YAML frontmatter (system, paths) to bucket files into systems. Declaring systems upfront lets /smith-index route files correctly from day one. This step is optional — skip if the project is small, exploratory, or doesn't yet have clear system boundaries. Systems can also be added later manually or via /smith-migrate-system-paths on a project that grew .specify/systems/ organically.

Ask the operator:

"Does this project have logical systems you'd like to declare upfront, so /smith-index can route files correctly from day one? (yes/skip)"

If the operator answers anything other than yes / y / Y, skip this step silently.

If yes, enter a per-system loop:

1. Ask for a short kebab-case name for the system (e.g. email-pipeline, command-center, auth). Re-prompt on blank input. Strip surrounding whitespace.
2. Compute the system id as system-NN-<short-name> where NN is the count of .specify/systems/system-*/ directories that exist so far in this run, plus one, zero-padded to two digits (01, 02, ...).
3. Ask for directory prefixes that belong to this system, one per line. Terminate the inner loop on a blank line. For each entry:
   • Strip surrounding whitespace.
   • Reject entries containing any glob character (*, ?, [, ], {, }, !) — v1 supports literal prefixes only. Print: "Glob characters not allowed in v1 — please enter a literal prefix." and re-prompt for the same slot.
   • Auto-append a trailing / if missing.
   • Collect into a list.
4. Copy the template: read ~/.claude/skills/smith/templates/system-spec-template.md, and write it to .specify/systems/<id>/spec.md with these substitutions in the YAML frontmatter block ONLY (do not touch the body):
   • system: system-<NN>-<short-kebab-name> → system: <id>
   • The single placeholder list item - <relative-prefix>/ is replaced by one - <prefix> line per collected entry, preserving two-space indentation. If the list is empty, emit paths: [] on a single line in place of the multi-line block.
   • status: draft stays as-is.
   • also_affects: [] stays as-is.
   • The body skeleton stays as-is — the user can flesh it out later (or A3 can be re-run, though A3 only adds paths, not body).
5. Ask whether to declare another system. If yes, loop back to step 1. If no, exit the loop.

Validation invariant (enforced by step 3): every prefix written into the paths: list contains no glob characters and ends with /. The resolver loads paths defensively, but this step is also the first line of defense.

After the loop completes, print a summary:

Scaffolded N system spec(s):
  - .specify/systems/system-01-<name>/spec.md (2 paths)
  - .specify/systems/system-02-<name>/spec.md (1 path)
You can edit the body of each spec at any time. Run `/smith-migrate-system-paths` to bulk-propose paths from existing prose.

If the operator types skip at any per-system prompt, skip that system (do not write a spec for it) and continue the outer "another system?" loop.

4.10 Clear Bootstrap Marker

The bootstrap marker created in Phase 4.1 has done its job — Phase 4 file generation is complete. Remove it via the shipped helper:

.specify/scripts/bash/clear-active-workflow.sh "bootstrap"

After this, the workflow-gate hook enforces normal discipline: subsequent file edits require one of the four top-level workflows (/smith-new, /smith-bugfix, /smith-debug, /smith-build) or /smith-finish to be active.

Phase 5: Verification & Report

After generating all files, present a summary:

## SpecKit Initialization Complete

### Files Created
| File | Status |
|------|--------|
| `CLAUDE.md` | Created / Updated |
| `.specify/memory/constitution.md` | Created |
| `.specify/templates/` (5 files) | Created |
| `.specify/scripts/bash/` (5 files) | Created |
| `.claude/commands/` (10 files) | Created |
| `.claude/agents/` (7 files) | Created |
| `.claude/settings.json` | Created / Skipped (exists) |
| `docs/sessions/` | Created (session chat logs) |
| `specs/questions/` | Created (pre-implementation question files) |
| `specs/init-intake.md` | Created (permanent project config decisions) |
| `.specify/systems/system-*/spec.md` (N files) | Created / Skipped (4.8) |

### Available Commands
| Command | Description |
|---------|-------------|
| `/smith.specify <description>` | Create feature spec |
| `/smith.clarify` | Clarify spec ambiguities |
| `/smith.plan` | Generate implementation plan |
| `/smith.tasks` | Generate task breakdown |
| `/smith.analyze` | Cross-artifact consistency check |
| `/smith.implement` | Execute tasks |
| `/smith.checklist` | Generate quality checklist |
| `/smith.constitution` | Update project constitution |

### Next Steps
1. Review `CLAUDE.md` and `.specify/memory/constitution.md` — adjust any generated content
2. Commit the generated files: `git add CLAUDE.md .specify/ .claude/ && git commit -m "chore: initialize SpecKit workflow"`
3. Start your first feature: `/smith.specify <your feature description>`

Important Notes

• Never modify files in the source project — this skill only reads from ~/.claude/skills/smith/ and writes to the current working directory
• Existing files: Always ask before overwriting. If CLAUDE.md exists, offer to merge or replace
• Idempotent: Running /smith.init again should detect existing setup and offer to update
• Minimal questions: Skip any question where codebase detection has high confidence. The goal is to confirm, not interrogate
• The generated CLAUDE.md and constitution are starting points — users should customize them for their specific needs