| name | forge-setup-project |
| description | Set up a project's meta-structure for agentic engineering — CLAUDE.md, AGENTS.md, docs/, README, and CHANGELOG. Use when starting a new project or retrofitting an existing codebase for the forge workflow. |
| disable-model-invocation | true |
| allowed-tools | Read, Write, Bash, Grep, Glob |
Set Up Project Meta-Structure
Set up a project's meta-structure for agentic engineering — the foundational files that make a codebase navigable and maintainable by both humans and AI agents.
Input
Optional: A path to the project root: $ARGUMENTS
If no argument is provided, use the current working directory.
Process
Step 1: Determine Project State
Scan the project root to classify what exists:
ls -la CLAUDE.md AGENTS.md README.md CHANGELOG.md .claude/settings.local.json 2>/dev/null
ls docs/ 2>/dev/null
ls -la
Classify the project into one of four states:
| State | Description | Behavior |
|---|
| Greenfield | No code, no meta files | Generate everything from user input |
| Existing, no meta | Has code but no CLAUDE.md/docs/ | Explore codebase first, then generate |
| Partial meta | Has some meta files (e.g., README exists) | Ask user how to handle each existing file |
| Full meta | All meta files already present | Offer to audit and update |
If existing meta files are found, use AskUserQuestion to ask:
- Which files to overwrite vs. keep
- Whether to merge content (for README.md especially)
Step 2: Explore the Codebase
Skip this step for greenfield projects.
For existing projects, perform a thorough exploration:
find . -maxdepth 3 -type f -not -path './.git/*' -not -path './node_modules/*' -not -path './vendor/*' -not -path './.next/*' -not -path './dist/*' -not -path './build/*' | head -100
ls package.json go.mod Cargo.toml pyproject.toml setup.py Gemfile build.gradle pom.xml mix.exs deno.json bunfig.toml composer.json 2>/dev/null
cat package.json 2>/dev/null | head -80
cat Makefile 2>/dev/null | head -80
cat Taskfile.yml 2>/dev/null | head -80
ls -la src/ app/ cmd/ lib/ main.* index.* 2>/dev/null
ls .github/workflows/*.yml .gitlab-ci.yml Jenkinsfile .circleci/config.yml 2>/dev/null
ls .eslintrc* biome.json .prettierrc* .golangci.yml rustfmt.toml .rubocop.yml pyproject.toml 2>/dev/null
ls jest.config* vitest.config* pytest.ini .mocharc* 2>/dev/null
ls Dockerfile docker-compose*.yml 2>/dev/null
ls docs/*.md *.md 2>/dev/null
Record what you discover:
- Language(s) and runtime
- Package manager
- Available scripts/make targets (build, test, lint, format, dev, etc.)
- Project structure pattern (monorepo, single app, CLI tool, library, etc.)
- Test framework
- CI/CD setup
- Entry points and architecture style
Step 3: Gather Project Information
Use AskUserQuestion to collect what can't be determined from code:
Always ask:
- One-line project description — "What does this project do in one sentence?"
- Core principles — "What are 3-5 rules that should always guide development in this project?" (Offer examples based on what you've seen in the codebase)
- Project story — Suggest a short metaphor or narrative that connects the project name to its purpose (2-3 sentences). This goes in the README header to give the project personality. Look for a metaphor grounded in what the name evokes — a physical object, a place, an action — and connect it to what the software actually does. Propose a suggestion and let the user refine it.
Ask only if ambiguous from code:
4. Confirm detected commands if multiple options exist (e.g., npm vs bun)
5. Confirm project structure if unclear (monorepo vs single app)
6. External dependencies — "Does this project depend on external services or APIs? List them with how to verify they're working (e.g., health-check endpoints, test commands)."
7. Debugging patterns — "When something breaks, what should be checked first?" (e.g., env/config files, specific log locations, common failure modes)
8. Preferred libraries — "Are there preferred libraries or patterns for common tasks that new code should follow?"
Never ask what's discoverable from code:
- Language, framework, test runner, linter — detect these automatically
- Available scripts — read package.json/Makefile
- Directory structure — explore it
Step 4: Generate CLAUDE.md
Create CLAUDE.md following this exact structure:
# <Project Name>
<one-line description from Step 3>
## Commands
<all discovered commands in a code block — install, dev, build, check, lint, format, typecheck, test, etc.>
## Documentation
| Document | Purpose |
|----------|---------|
| [Architecture](docs/architecture.md) | System design, data flow, package responsibilities |
| [Development](docs/development.md) | Prerequisites, setup, daily workflow |
| [Coding Guidelines](docs/coding-guidelines.md) | Code style, error handling, naming conventions |
| [Testing](docs/testing.md) | Test commands, conventions, patterns |
| [PR Workflow](docs/pr-workflow.md) | Commits, PRs, branch naming, review process |
<additional project-specific docs rows if applicable>
## Core Principles
<3-5 principles from Step 3, as a numbered list>
## Commits
Format: `<type>(<scope>): <description>`
Types: feat, fix, docs, refactor, test, chore, perf
<2-3 example commit messages using actual project scopes>
Optional sections: Include ## External Dependencies, ## Debugging, and ## Conventions only when the user provides relevant information in Step 3. Do not generate empty sections or invent content.
Critical rules:
- Commands block must contain ONLY commands that actually exist in the project
- Every command must be verified against package.json scripts, Makefile targets, or equivalent
- Use
<!-- TODO: verify this command --> if uncertain about a command
- Core principles must come from user input, not be invented
Step 5: Create AGENTS.md, .claude/settings.local.json, and .gitignore
ln -sf CLAUDE.md AGENTS.md
mkdir -p .claude
If .claude/settings.local.json already exists, merge the attribution keys into it. If it doesn't exist, create it:
{
"attribution": {
"commit": "",
"pr": ""
}
}
The empty attribution fields suppress Claude Code's default Co-Authored-By lines. This file is user-local and not committed — the "no attribution" rule is also documented in CLAUDE.md for enforcement.
Create .gitignore if it doesn't exist. If it already exists, ensure .claude/ is listed.
The .claude/ directory contains only user-specific settings (settings.local.json with tool permissions and attribution) and should not be committed.
if [ ! -f .gitignore ]; then
echo ".claude/" > .gitignore
else
grep -qxF '.claude/' .gitignore || echo '.claude/' >> .gitignore
fi
Add other standard entries based on the detected tech stack:
- Node/Bun:
node_modules/, dist/, .env, .env.local
- Go: binary name,
vendor/ (if not committed)
- Python:
__pycache__/, .venv/, *.pyc
- Rust:
target/
- General:
.DS_Store, *.log, coverage/
IMPORTANT: If .gitignore already exists, do NOT overwrite it. Only append missing entries.
Step 6: Generate /docs
Create the docs/ directory and generate five core documentation files plus any project-specific extras.
mkdir -p docs
Five core docs (always created):
docs/architecture.md
# Architecture
<system design overview — what the project is, how it's structured>
## Project Structure
<directory tree with annotations for key directories>
## Data Flow
<how data moves through the system — request lifecycle, processing pipeline, etc.>
## Package/Module Responsibilities
<table or list of key packages/modules and what they do>
<!-- TODO: Add diagrams if helpful -->
docs/development.md
# Development
## Prerequisites
<runtime, language version, tools needed — detected from project>
## Setup
<step-by-step from clone to running — based on actual project setup>
## Daily Workflow
<common development loop — start dev server, run tests, etc.>
## Available Commands
<full list of all scripts/make targets with descriptions>
docs/coding-guidelines.md
# Coding Guidelines
## Code Style
<detected style rules — formatter, linter config, import ordering, etc.>
## Error Handling
<project's error handling patterns — detected from code or marked TODO>
## Naming Conventions
<file naming, function naming, variable naming patterns detected>
## Documentation
<when to add comments, docstring conventions, etc.>
<!-- TODO: Add examples from the codebase -->
docs/testing.md
# Testing
## Running Tests
<exact test commands from project>
## Test Conventions
<test file location, naming, framework-specific patterns>
## Writing Tests
<patterns detected in existing tests — describe what you see>
## Coverage
<coverage commands if available, coverage expectations>
docs/pr-workflow.md
# PR Workflow
## Commit Conventions
Format: `<type>(<scope>): <description>`
Types: feat, fix, docs, refactor, test, chore, perf
### Examples
<2-3 examples using actual project scopes>
## Branch Naming
Format: `<type>/<issue-number>-<short-kebab-description>`
### Examples
<2-3 examples relevant to the project>
## PR Checklist
- [ ] Code follows project guidelines (see [Coding Guidelines](coding-guidelines.md))
- [ ] Tests added/updated (see [Testing](testing.md))
- [ ] Documentation updated (if applicable)
- [ ] CHANGELOG.md updated for user-facing changes
- [ ] Lint/format checks pass
- [ ] All tests pass
## Review Process
<project-specific review norms, or standard forge workflow>
Additional project-specific docs — offer based on detection:
| Detected Signal | Suggested Doc |
|---|
| REST/GraphQL routes, OpenAPI spec | docs/api-reference.md |
| Dockerfile, docker-compose | docs/deployment.md |
| CI workflow files | docs/ci.md |
| Auth middleware, security headers | docs/security.md |
| React/Vue/Svelte components | docs/frontend.md |
| Database migrations, ORM config | docs/database.md |
| Multiple packages/workspaces | docs/monorepo.md |
Use AskUserQuestion to offer detected extras:
Based on what I found in the codebase, I'd also suggest creating:
- docs/api-reference.md (detected REST routes in src/routes/)
- docs/deployment.md (detected Dockerfile and docker-compose.yml)
Which of these would you like me to create?
Content rules for all docs:
- Every sentence must be specific to THIS project — no generic filler
- Use
<!-- TODO: ... --> for anything that can't be determined from code
- Reference actual file paths, actual commands, actual patterns found in the codebase
- Keep each doc focused — if it's getting long, you're adding too much
Step 7: Generate README.md
If no README.md exists, create one following this structure:
<p align="center">
<strong><one-line project description></strong><br>
<short tagline — a punchy subtitle that complements the description>
</p>
<p align="center">
<a href="docs/architecture.md">Architecture</a> ·
<a href="docs/development.md">Development</a> ·
<a href="docs/coding-guidelines.md">Guidelines</a> ·
<a =>Testing
PR Workflow
---
---
See []() for full setup instructions.
| Document | Purpose |
|----------|---------|
| []() | System design and data flow |
| []() | Setup and daily workflow |
| []() | Code style and conventions |
| []() | Test commands and patterns |
| []() | Commits, PRs, and review process |
Create an issue:
Implement:
Self-review:
Address feedback:
Update changelog:
Header rules:
- If the project has a logo (
assets/logo.svg or similar), add a centered <img> above the tagline
- The story paragraph goes between two
--- dividers, right after the nav links
- Keep the story to 2-3 sentences — evocative, not verbose
If README.md already exists, use AskUserQuestion:
A README.md already exists. How would you like to proceed?
1. Replace it entirely with the new structure
2. Merge — keep existing content and add missing sections
3. Keep the existing README.md unchanged
Step 8: Generate CHANGELOG.md
If no CHANGELOG.md exists, create one with header only:
# Changelog
All notable user-facing changes to this project will be documented in this file.
Changes are grouped by release date and category. Only user-facing changes are included — internal refactors, test updates, and CI changes are omitted.
If the project has existing git history, use AskUserQuestion:
This project has existing git history. Would you like me to backfill the changelog from recent commits?
1. Yes — scan recent commits and create initial entries
2. No — start fresh from this point forward
If backfilling, follow the forge-update-changelog conventions (user-facing, plain language, no jargon).
Step 9: Commit
Stage all new and modified meta files and commit:
git add CLAUDE.md AGENTS.md .gitignore docs/ README.md CHANGELOG.md
git commit -m "docs: add agentic engineering meta-structure"
Do NOT commit if:
- The user asked for a dry run
- There are unrelated staged changes (unstage them first)
- Any generated file has unresolved questions (ask user first)
Step 10: Summary
Present a summary of everything that was created:
## Setup Complete
### Files Created
- CLAUDE.md — Project guide with commands, docs table, principles, commit conventions
- AGENTS.md — Symlink → CLAUDE.md
- .claude/settings.local.json — Attribution settings (user-local, not committed)
- .gitignore — Git ignore rules (includes .claude/)
- docs/architecture.md — System design and structure
- docs/development.md — Prerequisites, setup, workflow
- docs/coding-guidelines.md — Code style and conventions
- docs/testing.md — Test commands and patterns
- docs/pr-workflow.md — Commit, PR, and review conventions
<any additional docs created>
- README.md — Project overview and quick start
- CHANGELOG.md — Ready for entries
### Next Steps
1. Review each generated file and fill in any <!-- TODO --> markers
2. Use /forge-create-issue to plan your first piece of work
3. Use /forge-implement-issue to start implementing
Important Guidelines
Content Quality
- No generic boilerplate: Every sentence must be specific to the project. "This project uses React" is specific. "Follow best practices" is not.
- Placeholders over fiction: If you can't determine something from the code, use
<!-- TODO: describe X --> rather than making something up.
- Derived from exploration: Commands, structure, patterns — all must come from actual codebase analysis, not assumptions.
What NOT to Do
- Don't generate docs for a tech stack you didn't detect
- Don't invent project principles — ask the user
- Don't include attribution lines (Co-Authored-By, etc.) in any commits
- Don't include time estimates anywhere
- Don't add commands to CLAUDE.md that don't exist in the project
- Don't create AGENTS.md as a regular file — it must be a symlink
- Don't generate content that contradicts existing project configuration
Handling Edge Cases
- Monorepo: Create root-level meta files that reference sub-packages. Suggest per-package CLAUDE.md files as a follow-up.
- No tests yet: Create
docs/testing.md with <!-- TODO --> markers and recommend setting up a test framework.
- No CI yet: Skip
docs/ci.md but mention it as a follow-up in the summary.
- Multiple languages: Document all detected languages in CLAUDE.md commands section, grouped by language.
Related Skills
Next step: Use forge-create-issue to plan your first piece of work.
Full workflow: forge-setup-project → forge-create-issue → forge-implement-issue → forge-reflect-pr → forge-address-pr-feedback → forge-update-changelog
Example Usage
/forge-setup-project
/forge-setup-project /path/to/project