/project-init

Start a fresh, virgin project from scratch. This skill runs an interactive, dynamic onboarding dialog with the user, then scaffolds a complete, optimized directory & file structure and writes the onboarding results into the project's ops/ docs so that future LLM sessions (via the session-start trigger) can pick up the context automatically.

What this is for

• A completely empty folder that should become a new project.
• The user wants the LLM to ask the right questions rather than guess.
• All project areas must be covered: goal, stack, structure, tooling, docs, secrets/tokens, platform context.
• Output is a reusable, consistent baseline that the session-start routine can read back in every session.

PROTECTION RULE - never ~/.claude/

Read-only skill. Guard required if write mode added later.

What You Must Do When Invoked

During analysis, assign a confidence level to each finding: proven (confirmed by evidence), likely (strong signal, needs review), or suspected (weak signal).

If the user invoked /project-init -help or /project-init -h (no other args), print the contents of the ## Usage section verbatim and stop.

Otherwise follow the steps below in order. Do not skip steps.

Step 1 - Help check

If invoked with -help or -h, output the ## Usage section unchanged and stop.

Step 2 - Establish target directory

Ask the user for the absolute path of the new project (or use the current directory). Confirm it is empty or near-empty. If it already contains a project, stop and warn before overwriting anything.

Target directory: <path>
Existing files:   <list or "none">
Proceed? (yes/no)

Only continue after explicit confirmation.

PROTECTION RULE - never ~/.claude/: The target directory must under no circumstances be C:\Users\ostol\.claude\ (or its subdirectories). This path is sacred and must NEVER be modified by this skill. If the user proposes ~/.claude/ as target, abort immediately. The generator (scripts/init.ps1) blocks such paths technically by itself.

Step 3 - Run the interactive onboarding dialog

Ask the user the following areas ONE AT A TIME, in this order. Adapt follow-up questions to previous answers (dynamic, not a fixed form). Keep each question short and direct.

1. Goal - What is the project? One sentence. What problem does it solve?
2. Stack - Language/framework (Node/TS, Python, Go, Rust, ...). Confirm package manager. If unknown, propose a sensible default and let user accept/change.
3. Directory structure - Based on stack, propose a layout (src/, tests/, docs/, ops/). Confirm or adjust.
4. Tooling - Lint, formatter, tests, CI, git hooks, type-check. Confirm which to include.
5. Docs / Ops - Explain that ops/ holds manifest.md, tracking.md, sprints/. Confirm naming conventions.
6. Secrets / Tokens - Which external services/APIs? Where do secrets live (.env, vault)? NEVER log full secrets - only masked preview.
7. Platform context - Deployment target (local, cloud, serverless), platforms (web, cli, api, mobile).
8. Blockers / Open - Anything currently blocking or undecided?

After each answer, reflect it back in one line so the user can correct.

Step 4 - Generate the structure

Run the generator script. It reads the collected answers (you pass them as arguments / a JSON file) and creates the full tree.

See scripts/init.ps1 for the implementation. Invoke it like:

& "<SKILL_DIR>/scripts/init.ps1" -ProjectDir "<path>" -AnswersFile "<path/to/answers.json>"

Where answers.json is a JSON object you build from the dialog, e.g.:

{
  "name": "my-app",
  "goal": "Track personal finances.",
  "stack": "node-ts",
  "pkgManager": "npm",
  "layout": "src/tests/docs/ops",
  "tooling": ["eslint", "prettier", "vitest", "github-actions"],
  "docs": ["manifest", "tracking", "sprints"],
  "secrets": [".env (gitignored)"],
  "platform": ["web", "api"],
  "blockers": "none"
}

The script:

• Creates the directory tree.
• Writes CLAUDE.md (project rules and configuration).
• Writes ops/manifest.md (goal + scope from answers).
• Writes ops/tracking.md (status template, blocker field).
• Writes ops/sprints/.gitkeep + ops/sprints/README.md.
• Writes .gitignore (node_modules, dist, .env, etc. - stack-aware).
• Writes README.md (project title + goal stub).
• Scaffolds a minimal entry file appropriate to the stack (only a stub, no logic).

Step 5 - Finalize & hand off

Print a concise summary of what was created:

Project <name> initialized in <path>
  Structure: <tree in 5-8 lines>
  Config:   CLAUDE.md created (project rules)
  Ops:       manifest.md, tracking.md, sprints/
  Secrets:   <masked preview or "none">
  Next step: type "weiter" to start session start routine.

Tell the user: from now on, typing "weiter" in this project triggers the Session-Start-Routine that reads manifest/tracking/sprints automatically.

Usage

/project-init                 # interactive onboarding in current directory
/project-init <path>          # interactive onboarding in <path>
/project-init -help          # show this usage block and stop