init-project

name	init-project
description	Use when creating a new project. Bare AI shell by default; interactive menu adds docs / backend-infra / agents on demand.

Init Project

Type: Engineer (light)

Overview

Bootstrap a new project with the minimal Claude Code shell (CLAUDE.md, .gitignore, .claude/settings.json). In interactive mode an opt-in menu adds documentation, backend infrastructure, and agent stubs — only the modules you actually tick. Everything runs via focused bash modules under scripts/modules/; no LLM cost for the mechanical scaffold work.

Process

Step 0: Pre-flight (Opus main)

All Q&A happens here before any bash is executed.

1. Verify working directory. Run ls -A on the target directory (default: cwd). If it is non-empty — anything beyond .git/ alone — ask via AskUserQuestion:

   "The current directory is not empty. Contents: <list>. Continue and add project structure here, or abort?"

   • Abort → exit cleanly. Do not proceed.
   • Continue → proceed. Modules skip files that already exist.

2. AskUserQuestion — project name (free text):

   "What is the project called?"

3. AskUserQuestion — mode (single-select):

   "Scaffold mode?"

   • Options: minimal | interactive
   • minimal — only core module runs (bare AI shell).
   • interactive — core runs first, then the module menu (Step 2) is presented.

4. Resolve plugin root.

   • Glob ~/.claude/plugins/cache/vladyslav-marketplace/vladyslav/*/scripts/modules/core.sh.
   • Take the matching path and go three levels up (strip scripts/modules/core.sh) to get <root>.
   • Fall back to /Volumes/DevSSD/Development/vladyslav-skills if Glob returns nothing (development clone).
   • Verify <root>/skills/init-project/assets/ exists. If missing, warn the user and continue — assets are optional for core operation.

Capture: <pwd>, <name>, <mode>, <root>.

---

Step 1: Core (always)

Execute via Bash tool:

<root>/scripts/modules/core.sh \
    --pwd <pwd> \
    --plugin-root <root> \
    --name <name>

The script emits a single line of JSON on stdout:

{
  "status": "success" | "partial" | "error",
  "files_written": ["<path>", ...],
  "files_skipped": ["<path>", ...],
  "warnings": ["<msg>", ...],
  "error": "<msg if status=error>"
}

Capture this JSON. If status is error, surface the error and stop — do not proceed to Step 2.

---

Step 2: Module menu (interactive mode only)

Skip entirely if <mode> is minimal.

Present three grouped AskUserQuestion multi-selects, each with all options unticked by default. Run only the modules the user ticks. Unticked items are silently skipped — "don't know → don't create".

Group A — Docs

"Which documentation scaffolds do you want?" (all unticked by default)

Option	Module call
prd + planning	docs.sh --pwd <pwd> --plugin-root <root>
design system	design-system.sh --pwd <pwd> --plugin-root <root> --name <name>
architecture	architecture.sh --pwd <pwd> --plugin-root <root>

Group B — Backend infra

"Which backend infrastructure do you want?" (all unticked by default)

Option	Module call
docker	docker.sh --pwd <pwd> --plugin-root <root>
postgres	postgres.sh --pwd <pwd> --plugin-root <root>
redis	redis.sh --pwd <pwd> --plugin-root <root>
alembic	alembic.sh --pwd <pwd> --plugin-root <root>
backend skeleton	backend-skeleton.sh --pwd <pwd> --plugin-root <root>

Group C — Agents

"Which Claude Code agents do you want installed?" (all unticked by default)

Options: architect-reviewer, backend-engineer, qa-reviewer, release-manager

Collect the ticked names into a comma-separated string, then run:

<root>/scripts/modules/agents.sh \
    --pwd <pwd> \
    --plugin-root <root> \
    --agents "<csv>"

If no agents are ticked, skip the agents.sh call entirely.

Each module call emits the same JSON shape as core. Capture all outputs.

---

Step 3: Roadmap gate

After modules complete, ask via AskUserQuestion:

"Які ключові фічі плануєш в цьому проекті? Розіб'ємо на MVP-фази в ROADMAP.md"

• If the user provides features → generate ROADMAP.md at <pwd>/ROADMAP.md using the phase structure below, then commit the file to git.
• If the user skips ("потім", "не знаю", empty answer, or declines) → do not create the file. Continue normally.

This step is non-blocking — init-project completes regardless of the answer.

ROADMAP.md phase structure:

# Roadmap: <name>

> Created: YYYY-MM-DD

## Phase 1: <Name>
**Done when:** <one sentence criteria>

- [ ] Task 1
- [ ] Task 2
- [ ] Task 3

## Phase 2: <Name>
**Done when:** <one sentence criteria>

- [ ] Task 1
- [ ] Task 2

Phases represent MVP milestones — what ships in the first release vs. what comes after.

---

Step 4: Summary + smart Next

Merge all module JSON outputs. Render to the user:

✓ init-project complete
  Project: <name> in <pwd>
  Mode: <minimal|interactive>
  Files written: <total count across all modules>
  Skipped (already existed): <list, omit section if empty>
  Warnings: <list, omit section if empty>
  ROADMAP.md: <created | skipped>

Then append a context-aware Next: line — pick the first that applies:

Condition	Next line
No docs module was ticked	Next: docs народяться за потреби — discover / write-user-stories / write-test-docs.
Any backend-infra module was ticked	Next: /vladyslav:add-feature щоб почати фічу.
minimal mode	Next: повернись у interactive за потреби, або одразу /vladyslav:add-feature.

The skill never auto-runs the next skill. The Next line is informational only.

status: error from any module: surface the error and which module failed. List files written before the failure. Do not retry automatically.

---

Why this is a Light Engineer skill

• No subagent dispatch. Each module is a deterministic bash script. No LLM reasoning is needed for file creation.
• Modular and opt-in. Only requested modules run, so a minimal project stays minimal.
• Idempotent. Modules skip files that already exist and report them in files_skipped.
• ~1 second, 0 LLM tokens per module. Q&A in Step 0 is the only model-driven part.

---

Output allowlist

Paths the modules may write (relative to <pwd>). The modules never modify pre-existing files except where noted (append operations on docker-compose.yml and .env).

core.sh (always):

.gitignore
.claude/settings.json
CLAUDE.md
.remember/now.md

docs.sh:

docs/product/prd.md
docs/plans/tasks.md
docs/plans/backlog-next.md

design-system.sh:

docs/design/system.md

architecture.sh:

docs/architecture/system.md

docker.sh:

Dockerfile
docker-compose.yml
docs/operations/docker.md

postgres.sh:

docker-compose.yml              (appends postgres service block)
.env                            (appends DATABASE_URL)

redis.sh:

docker-compose.yml              (appends redis service block)
.env                            (appends REDIS_URL)

alembic.sh:

alembic.ini
migrations/env.py
migrations/versions/.gitkeep
migrations/README

backend-skeleton.sh:

requirements.txt
src/__init__.py
src/main.py

agents.sh:

.claude/agents/<name>.md        (one per chosen agent)

Roadmap gate (Step 3):

ROADMAP.md