name: init-project description: Initialize a complete project from scratch. Asks all necessary questions, designs the architecture, generates a filled CLAUDE.md from the global template, writes a cross-platform README with setup instructions, installs dependencies, and delivers a working first version covering all v1 features. argument-hint: disable-model-invocation: true

allowed-tools: Read, Write, Edit, Bash, Grep, Glob

ORCHESTRATOR: INIT PROJECT

AGENTS LOADED

Load and follow strictly:

.claude/agents/interviewer.md
.claude/agents/analyzer.md
.claude/agents/designer.md
.claude/agents/scaffolder.md
.claude/agents/reviewer.md
.claude/agents/tester.md

INITIAL REQUEST

$ARGUMENTS

WORKFLOW

STEP 1 — INTERVIEWER

Run the INTERVIEWER agent.

Using the initial request above as a starting point:

Identify which information is already clearly provided.
Ask only what is genuinely missing or ambiguous.
Present all remaining questions in a single structured block.

MANDATORY STOP — do not continue until the user has answered.

After receiving answers, produce the PROJECT BRIEF as defined in interviewer.md. The PROJECT BRIEF is the single source of truth for all subsequent steps.

STEP 2 — ANALYZER

Run the ANALYZER agent on the PROJECT BRIEF.

Focus on:

Existing repo or code to integrate (if any)
Stack constraints and compatibility issues
Infrastructure or environment constraints
Risks that could affect the design
Any open decisions listed in the PROJECT BRIEF

Output an ANALYSIS REPORT.

STEP 3 — DESIGNER

Run the DESIGNER agent using the PROJECT BRIEF and ANALYSIS REPORT.

Produce a complete DESIGN covering:

Final tech stack with exact versions
Complete folder structure (full tree)
Module responsibilities and data flow
Key interfaces and data models
Config and tooling setup (lint, format, CI)
Test strategy and tooling
Any open decisions resolved with justification
Prerequisites list (what must be installed on the dev machine)

STEP 4 — VALIDATION GATE

MANDATORY STOP — present the following to the user and wait for explicit approval before proceeding.

================================================================
INIT PROJECT — VALIDATION GATE
================================================================

PROJECT BRIEF SUMMARY
---------------------
<3–5 line recap of what will be built>

STACK
-----
<finalized stack with versions>

PREREQUISITES TO INSTALL
-------------------------
<list of tools / runtimes / services with versions>

FOLDER STRUCTURE
----------------
<full tree from the DESIGN>

V1 FEATURES TO IMPLEMENT
-------------------------
<numbered list from PROJECT BRIEF>

CONVENTIONS
-----------
<naming, doc style, test strategy>

EXCEPTIONS TO GLOBAL RULES
---------------------------
<list or "none">

================================================================
Approve this plan? (yes / request changes)
================================================================

IF user requests changes:

Return to STEP 3 (DESIGNER) with the feedback.
Repeat STEP 4.

IF approved → proceed to STEP 5.

STEP 5 — SCAFFOLDER

Run the SCAFFOLDER agent with:

The full PROJECT BRIEF
The approved DESIGN
Reference to ~/.claude/templates/project-CLAUDE.md
Reference to ~/.claude/CLAUDE.md

The SCAFFOLDER will, in order:

Generate CLAUDE.md — fill the global template with real content from the PROJECT BRIEF. No placeholders. No examples.
Generate README.md — cross-platform setup instructions (Windows / Linux / macOS) covering:
- All prerequisites with exact versions and install commands
- Step-by-step installation
- How to run in development and production
- How to run tests
- Environment configuration
Generate Claude Code settings — create .claude/ with:

a. .claude/settings.json — from ~/.claude/templates/settings/settings.json.

  Adapt the `allow` rules to the actual project stack:
  - Keep only the tool blocks relevant to this stack
  - Add any stack-specific commands not already in the template
  - Add project-specific `ask` rules (deploy targets, DB commands)
  - Leave `deny` empty — global deny rules live in `~/.claude/settings.json`

b. .claudeignore — from ~/.claude/templates/settings/.claudeignore.

  Extend with project-specific exclusions:
  - Stack-specific build artifacts not already covered
  - Sensitive file patterns specific to this project
  - Directories identified in the DESIGN as generated or cache

c. After creating these files, print:

  ```
  ⚙️  SETTINGS SETUP
  .claude/settings.json  created — project-level permissions
  .claudeignore          created — file exclusions for Claude

  Manual step required:
  Copy ~/.claude/templates/settings/settings.local.json
  to .claude/settings.local.json and add it to .gitignore.
  This file is personal and must not be committed.
  ```

Scaffold structure — create every folder and file from the DESIGN with real content.
Implement v1 features — real working code for every feature in the PROJECT BRIEF. No stubs. No TODOs.
Write initial tests — at minimum one happy path and one edge case per module.
Install and build — actually run the install command, build, and test suite. Fix any failures before reporting.

STEP 6 — REVIEWER

Run the REVIEWER agent on the scaffolded project.

Review scope:

Structure coherence vs approved DESIGN
Code quality vs CLAUDE.md conventions and global rules
Security issues in initial implementation
README completeness and accuracy
Generated CLAUDE.md completeness (no placeholder left)
Missing or incomplete v1 features

STEP 7 — FIX LOOP

Maximum 3 iterations.

IF CRITICAL issues found:

Run SCAFFOLDER with the list of issues to fix.
Run REVIEWER again.
Increment iteration counter.

IF counter > 3:

STOP.
Present blocking issues to the user and ask how to proceed.

IF only IMPORTANT or MINOR issues:

Proceed to STEP 8.
List all remaining issues in the final output.

STEP 8 — TESTER

Run the TESTER agent on the scaffolded project.

Produce:

Verification that existing tests cover v1 features
Additional test cases for identified edge cases
Regression risk assessment
Confirmation that the test command in CLAUDE.md is correct

RULES

Never skip the INTERVIEWER step.
Never start design or implementation with unanswered questions.
Never implement without explicit user approval of the DESIGN.
The generated CLAUDE.md must be complete — no placeholders.
The README must work on Windows, Linux, and macOS.
The first version must install, build, and run. A broken scaffold is not acceptable output.
Keep agents isolated in their responsibilities.

FINAL OUTPUT

================================================================
PROJECT INITIALIZED: <project name>
================================================================

LOCATION         : <project root path>
STACK            : <finalized stack>
INSTALL          : ✅ / ❌ <error>
BUILD            : ✅ / ❌ <error>
TESTS            : ✅ <N> passing / ❌ <detail>

V1 FEATURES
-----------
✅ <feature>
✅ <feature>
⚠️ <feature> — partial: <reason>

REMAINING ISSUES
----------------
<IMPORTANT and MINOR issues from reviewer, or "none">

QUICK START
-----------
<exact commands to get the project running right now>

NEXT STEPS
----------
1. <recommended first action>
2. <recommended second action>

CLAUDE.md        : ✅ complete
README.md        : ✅ Windows / Linux / macOS
SETTINGS         : ✅ .claude/settings.json + .claudeignore generated
================================================================

SKILL.md 7.8 KB 히스토리 Raw