Init project

Init Project

Purpose

Bootstrap a new project with the arthus-harness baseline (3 protection layers: process / technical / emotional). Runs the generator (npx create-arthus-harness) after a brief interview captures the inputs the generator needs.

This is the on-ramp skill — invoked once per project, then never again. After this runs, the harness is installed and the user's normal Claude Code workflow takes over.

When to use

• User says "init this project", "bootstrap with arthus-harness", or "set up the harness here".
• User opens a fresh repo (or a soon-to-be-fresh one) and wants protection layers in place before writing code.
• User is starting a new client project / MVP / side project and wants the same disciplined Claude Code setup.

When NOT to use

• The repo already has .claude/ and .arthus-harness.json — that means the harness is installed; the user wants harness-doctor (drift check) or arthus-harness sync (update) instead.
• The user wants to author a new harness skill / agent / hook — that's regular editing, not bootstrap.
• The user is asking how arthus-harness works conceptually — answer the question, don't run the generator.

Guard — refuse if project already initialized

Before doing anything else, check:

1. Does package.json exist in the cwd?
2. Does .claude/settings.json exist in the cwd?
3. Does .arthus-harness.json exist in the cwd?

If any of those exist, STOP. Tell the user:

This directory already has a project. Running init-project would overwrite or conflict. If you want to:

• Check drift vs current harness version → use harness-doctor skill.
• Update to a newer harness version → run npx arthus-harness sync from this directory.
• Start fresh anyway → cd to an empty parent directory and re-invoke me.

Do NOT proceed past the guard without explicit "yes, overwrite, I know what I'm doing" from the user.

Steps

1. Run the guard (above) — verify: cwd is empty enough to scaffold safely. Exit with the message above if not.

2. Interview the user — ask only the minimum, in order. Wait for each answer before continuing:

   a. Project name — kebab-case slug. Example: my-saas, client-mvp. Verify: matches ^[a-z][a-z0-9-]*$.

   b. One-line description — what the project does. Verify: <= 120 chars.

   c. Stack plugins — list available plugins (the generator prints them with --list-plugins). Ask which apply. Common: supabase, design-system, e2e-playwright, i18n-pt-br, payment-asaas. Verify: array of valid plugin names.

   d. Experience principles strategy — three options:

   • A (literal) — copy the pre-filled 4 sensations as-is, user adapts later.
   • C (framework) — empty template, user defines their own.
   • both — both, user picks per-doc.

   Verify: one of A / C / both.

3. Confirm with the user — print the chosen inputs back. Wait for "yes" / "go ahead" / "do it". Do NOT proceed without confirmation.

4. Run the generator — invoke:

   npx create-arthus-harness@latest <projectName> \
     --description "<description>" \
     --plugins <comma-separated> \
     --principles <A|C|both> \
     --in-place \
     --no-prompt
   

   --in-place tells the generator to scaffold INTO the current cwd (not create a subdirectory). --no-prompt skips the generator's own interactive prompts since we already gathered inputs.

   Verify: command exits 0; .claude/, Docs/, MISSION.md, CLAUDE.md, .arthus-harness.json exist after.

5. Print the next steps — show the user:

   Harness installed. Three things to do now:
     1. Read MISSION.md and adapt it to this project's invariants (~15 min).
     2. Read Docs/produto/principios-de-experiencia.md and define your sensations.
     3. Make your first commit — pre-commit hooks should fire.
   
   Then run your first feature task. Don't keep tuning the harness.
   

Examples

Example 1 — fresh client project

Input: "init this project"

Output:

• Skill asks 4 questions, confirms answers, runs npx create-arthus-harness client-saas --description "B2B SaaS for X" --plugins supabase,design-system --principles A --in-place --no-prompt.
• Reports installed files; prints next-steps message.

Example 2 — already initialized

Input: "init this project" (in a dir that already has .arthus-harness.json)

Output: Refuses with the guard message; suggests harness-doctor instead.

Escalation

• If npx create-arthus-harness is not found / fails → tell the user the package may not be published yet; they can git clone arthus-harness and run the generator from a local checkout.
• If the user wants flags not exposed by --no-prompt → drop the flag and let them go through the generator's own interactive prompts.
• Never: run the generator without confirmation; never overwrite an existing .claude/.

References

• Generator source: arthus-harness/src/
• Architecture: arthus-harness/PLAN.md
• Sister skill (post-install): harness-doctor