Scaffold Init

---

name: scaffold-init description: > Initialize an AI-native development workspace with spec-driven workflow infrastructure. Use when starting a new project, setting up Claude Code on a codebase for the first time, or when the user says scaffold, initialize, set up AI workflow, onboard this repo, or similar. Do not use for adding an individual skill or config to an already- scaffolded project — that is handled by the relevant tier skill directly. user-invocable: true

Spec 001 is fully implemented: greenfield-scaffold, doc-content, signal-detection, deferred-decisions, and Q&A wizard. See docs/specs/001-scaffold-init/spec.md.

What this skill does

Generates an AI-native development workspace by copying templates from ${CLAUDE_PLUGIN_ROOT}/templates/ into a target directory. Detects project signals from the filesystem (LLM/agent files, CI, tests, team), runs an optional Q&A flow to let the user override those signals, and selects tiers accordingly. Tier 0 always installs; Tier 1 installs when test signals are present; Tier 2 is offered (not auto-installed) when LLM/agent signals are present.

How to use

1. Determine the target directory. Default: the current working directory. If you're unsure, ask the user once before scaffolding.
2. Check if the target already has a scaffold.json or docs/specs/ — if so, the project is already scaffolded. Stop and tell the user rather than overwriting.
3. Run the Q&A flow (see next section). Collect answers as flag values.
4. Invoke the wizard with the collected flags:

   python3 "${CLAUDE_PLUGIN_ROOT}/skills/scaffold-init/scaffold.py" \
     [--runtime <name>] [--team|--solo] [--has-ci|--no-ci] \
     [--has-tests|--no-tests] [--plans-ai|--no-ai] \
     [--in-repo] \
     <target-dir>
   

   With no machinery flag the wizard scaffolds plugin mode — the lean default (docs + primer only; jig runs from the installed plugin). Pass --in-repo only when the sixth question is answered "yes".
5. Read the wizard's stdout summary and report back to the user. List the files that were created and the immediate next steps.

Codex custom-agent install

Codex scaffold mode (--host codex) writes project-local custom agents as TOML under .codex/agents/. For Codex plugin users who want jig's role agents globally available, run the explicit post-install helper:

python3 "${CLAUDE_PLUGIN_ROOT}/skills/scaffold-init/scaffold.py" --install-codex-agents

The default destination is ~/.codex/agents. Use --codex-agents-dir <dir> to target a different Codex agents directory. The helper refuses to overwrite user-owned jig-*.toml files unless --force is passed.

Q&A flow (slice 001-05)

Ask each question in order. Each question is independently skippable — if the user says "skip", "I don't know", "unsure", or similar, do not pass the flag (the wizard's filesystem inference handles it).

1. Runtime/language — "What runtime or language is this project? (e.g. Python, TypeScript, Go, Rust, mixed, unsure)" → --runtime <name> if answered; omit if skipped or unsure.
2. Team context — "Solo project or team setting?" → --team for team, --solo for solo; omit if skipped (uses git-author detection).
3. Existing CI — "Does the project already have CI configured?" → --has-ci for yes, --no-ci for no; omit if skipped.
4. Existing tests — "Does the project already have a test suite?" → --has-tests for yes, --no-tests for no; omit if skipped. This affects whether tier-1 (tdd-loop and friends) is auto-installed.
5. LLM/agent work planned — "Will this project involve LLM or agent development?" → --plans-ai for yes, --no-ai for no; omit if skipped. This affects whether tier-2 is offered.
6. Machinery vs. plugin — "Will this project ever run jig where the plugin may NOT be installed — CI, cloud agents, or teammates without jig? (yes → copy jig's machinery into the repo; no → use the installed plugin)" → --in-repo for yes; omit for no (plugin mode is the default). Skipping selects plugin mode. This is the one architectural question in the flow (it decides repo topology — what gets committed), so it won't appear in the project's own docs. Default to plugin mode (lean repo, jig updates flow from the plugin); choose --in-repo only for the self-contained cases above. See ADR-0041.

Skipping every question is the legitimate "pure inference" mode (slice 001-03 behavior) — the wizard infers from filesystem signals alone, and defaults to plugin mode. Do not invent answers when the user is unsure.

Output

After running, the target directory contains (plugin mode — the default):

• CLAUDE.md (with Hot Cache section, project name substituted)
• docs/ (architecture, workflow, conventions, refinement-todo, inbox, memory/, specs/, decisions/)
• .claude/hooks/ (empty — project-specific gates can go here)
• a project-scoped permissions file, on hosts that provide one — seeded with the ADR-0013 destructive-command deny floor (git push --force, git reset --hard, rm -rf). Hosts with no project-scoped permission surface get no such file and no deny floor.
• .gitignore (secret-ignore floor)
• scaffold.json (install-state manifest; scaffold_mode: "plugin-only")

For what your project actually received: scaffold.json records the mode and host, and your host's own settings file (if it has one) carries the deny floor. (This section is deliberately host-neutral — it is machine-translated per host.)

In plugin mode jig's skills, agents, and hooks stay under the installed plugin and run from the plugin root — no machinery is copied into the repo. The one exception is the permissions file above: permissions.deny lives in the project's own settings and no plugin mechanism can inject it, so the scaffold writes it in both modes wherever the host supports it (ADR-0041 OQ1). It carries no hook registrations.

The wizard's stdout summary states the mode and why.

With --in-repo, the target additionally gets a self-contained copy of jig's skills, agents, hook scripts, and templates under the host's runtime directory, plus the host's own hook-registration file (scaffold_mode: "in-repo"). Choose it only for CI, cloud agents, or teammates without jig installed.

Every scaffolded doc carries Status: Draft (wizard-generated). docs/memory/people.md is NOT created (solo-project default — team detection is slice 001-03).

Immediate next steps to surface to the user

After scaffolding succeeds, tell the user:

1. Open CLAUDE.md and fill in the Hot Cache section with project-specific terms.
2. Open docs/refinement-todo.md to see what was deferred.
3. The first spec to write is in docs/specs/ — use /jig:spec-workflow (when implemented) or write docs/specs/001-<feature>/spec.md by hand.
4. Note: docs/conventions.md is now gated. Edits require JIG_CONVENTIONS_APPROVED=1 in the shell session.

Constraints

• Do not invoke this skill in a directory that is already scaffolded (has scaffold.json).
• Do not overwrite an existing CLAUDE.md without explicit user confirmation.
• The wizard is deterministic — do not edit the generated files yourself before reporting back. The user should see exactly what scaffold.py produced.

Gotchas

• The spec-gate hook for docs/conventions.md activates AFTER scaffold-init completes. It cannot gate its own creation (bootstrap paradox — documented and intentional).
• templates/CLAUDE.md.template is the source template; do NOT use the jig repo's own CLAUDE.md as a template — the two diverge over time.
• ${CLAUDE_PLUGIN_ROOT} is the right env var inside the plugin. Don't confuse it with $CLAUDE_PROJECT_DIR (which is the target project's root after install).
• Signal detection (existing CI, LLM/agent files, team size) is deferred to slice 001-03. Until then, the wizard installs default tiers regardless of project context.
• scaffold-init refuses if the target looks spec-driven but lacks scaffold.json. Slice 008-05 introduced a second pre-flight check: if ≥3 of the four migrate triggers (docs/specs//docs/slices/, docs/decisions//docs/adrs/, docs/workflow.md, docs/architecture.md) are present without a scaffold.json, the helper raises LooksAlreadySpecDrivenError (exit 2) and points at /jig:migrate. Pass --force to scaffold over the existing tree anyway (NOT recommended — overwrites docs).