Setup

Big Emotion Project Setup

Every Big Emotion project ships the same toolset. This skill installs it, module by module, on any repo — or audits an existing repo against it. The full rationale lives in SPEC.md at the root of the agent-atelier repository.

Reference paths above are relative to ${CLAUDE_PLUGIN_ROOT}/skills/setup/. Templates live in ${CLAUDE_PLUGIN_ROOT}/skills/setup/templates/; the placeholder registry is templates/params.json.

Operating rules

• Never write before the plan is confirmed. Steps 1–3 are read-only; a single explicit confirmation covers the whole install plan.
• Idempotent. On a compliant repo, report "compliant" and change nothing. Never overwrite a project's existing customization of a module without flagging it as drift and asking.
• Modules are independent. Install only what the user selected; missing prerequisites between modules (M5 needs M4's board; M3's spec skill needs M4's config) are declared in the plan, not silently added.
• Everything written is in English (code, comments, docs). User-facing trigger phrases inside skills may stay French.
• Adapt, don't force. Templates ship in pnpm/TypeScript form; adapt commands to the interview answers (package manager, toolchain layout, no-TS repos) per each module's reference doc.
• Secret names, never values. The skill documents where every secret lives, what it is called, and how to obtain it — it never writes, echoes, or stores a secret value (not in rendered files, not in conversation output, not in this plugin). Values are entered by a human directly at their destination (GitHub secret, VPS .env, provider portal). This plugin is public and ships no infrastructure coordinates — no hostnames, IPs, SSH ports, account handles or resource names. Collect them from the user at interview time and never write them back into the plugin.

Step 1 — Interview

Read templates/params.json — it is the authoritative parameter list. Collect only what the selected modules need; propose defaults detected from the repo (package.json name → project_slug, git remote → org/repo, existing branches → branch model). For Atlassian ids, prefer discovery via Atlassian MCP over asking (see references/m4-atlassian.md).

Ask which modules to install when the user hasn't said; default to all seven for a new project. M7's coordinates (VPS host, SSH port and user, app hostname, deploy path, mail sender) ship with no defaults by design — ask for them, or let the user point you at their own private ops notes outside this repo.

Step 2 — Gap analysis (read-only)

For each selected module, classify missing / present / drifted with evidence:

• M1: .github/workflows/ci.yml exists? Contains the core gates (secret scan, lint, typecheck, format, test, build)? Partial = drifted, list absent gates.
• M2: .husky/ with pre-commit + commit-msg? commitlint.config.*, lint-staged.config.*, prepare: husky script?
• M3: .claude/skills/<slug>-{release,audit,spec,ticket,bootstrap-confluence}/? Also check ~/.claude/skills/ for name collisions — personal skills shadow project skills (warn, recommend deleting stale personal copies).
• M4: docs/confluence-spec/config.json complete? docs/.confluence-bootstrap-complete sentinel? docs/templates/jira-ticket-template.md? Jira board reachable via MCP with the five pipeline columns?
• M5: ferry.config.json|yaml + .github/workflows/ferry-router.yml (router model — five per-role ferry-*.yml workflows = drifted: legacy model, flag for migration but do not auto-migrate) + ferry-reconcile.yml + ferry-cost-daily.yml? Secrets/vars present (gh secret list, gh variable list)?
• M6: integration branch exists, default branch protected, deploy-production.yml triggered on tag v*?
• M7: deploy artifacts present (Dockerfile, deploy/docker-compose.yml, deploy/env.template — or Azure deploy workflows)? GitHub production environment with the DEPLOY_* variables + DEPLOY_SSH_KEY secret (gh api repos/{owner}/{repo}/environments)? Transactional mail wired per the M365 pattern where the app sends email?

Output one table: module · status · evidence · what install would do.

Step 3 — Plan + confirmation

Per-module install plan (files to write, wizards to run, manual steps the user must do in Jira/Confluence UIs, settings to change via gh). State what will NOT be touched. Wait for one explicit go.

Step 4 — Install

Work module by module in order M6 → M1 → M2 → M4 → M3 → M5 → M7 (branch model first — CI triggers reference branches; Ferry needs board, branches, and CI; infra last — its deploy workflow slots into M6's release flow). For each module follow its reference doc; render templates by replacing {{param}} tokens with interview values; # PROJECT-SPECIFIC: / <!-- PROJECT-SPECIFIC: --> markers in templates show where the project adds its own content — surface them to the user, never delete silently.

Manual Atlassian/GitHub-settings steps: guide the user through them interactively and verify each via MCP or gh before moving on.

Step 5 — Verify

• M2: make a scratch commit exercising the hooks (then reset) — both hooks must fire; a bad message must be rejected.
• M1: push a branch or open a draft PR to see the gates run, or gh workflow run where applicable.
• M3: each installed skill's preconditions pass (dry read-only run).
• M4: MCP reads of config ids succeed; sentinel present after bootstrap.
• M5: npx -p @big-emotion/ferry ferry-doctor; then a smoke Story through Refinement.
• Output a final checklist: done / remaining-manual / deferred.

Failure modes

• Repo already has a conflicting tool (e.g. lefthook, commitizen): stop, present the conflict, never replace without an explicit decision.
• Legacy Ferry (five per-role workflows) detected: report as drift with a migration pointer (references/m5-ferry.md); migration is its own task, not part of install.
• Atlassian MCP unavailable: M4/M3-spec install degrades to fully manual guidance; say so up front.
• User lacks org/repo admin rights (secrets, branch protection): emit the exact gh commands + settings paths for someone who has them, and mark those steps remaining.

Out of scope

Auto-creating Jira projects/boards (guided manual steps only) · migrating legacy-Ferry repos · website-specific gates (design tokens, Figma parity, Prismic checks — see the optional-extensions catalog in references/m1-ci.md) · anything requiring credentials the user hasn't provided.