SuperAPI AI Directory
Partner with us中文AI newsFavorites
Skillsincu-way-init

incu-way-init

Coding & development

View repository

Updated 10-15 22:37

About this skill

Repository Initialization Process

A gate-driven workflow for onboarding a repository into incu-way. It prepares the ground the other skills assume: a filled CLAUDE.md, a PRD.md, the docs/ tree, the develop/main branch model, and the .ways/ directory. For an existing application (brownfield — the common case) it also drives the architectural and functional documentation so that humans and the incu-way flows actually understand what the repo contains.

This is a bootstrap flow, not a work-item flow. It does not create a .ways/state.json (that model is for feature / bug / security work — see the ways/v1alpha1 state contract, shipped as the always-on state-contract rule). Instead, this skill creates the .ways/ directory the work-item flows later populate.

Repo eligibility check (before anything else)

Before Phase 0, check ~/.ways/config.yaml's blacklist (format: incu-base's global-config rule, installed always-on) for the current repo. If it matches, tell the user this repo is blacklisted from incu-way and ask whether to proceed anyway (a one-off exception) or stop — do not run any phase until they answer. Missing file, or no match: proceed normally.

Phase 0 — Orientation and mode detection

Goal: Decide whether this is a greenfield or a brownfield repository, and whether incu-way is already (partially) set up. Do not write anything yet.

Detection checklist

Run a read-only survey of the repo root:

git -C . rev-parse --is-inside-work-tree 2>/dev/null   # is this a git repo at all?
git log --oneline -1 2>/dev/null                        # any commits yet?
git branch -a                                           # does develop/main exist?
ls -A                                                    # what is at the root?

Then look for the signals below.

  1. Application code present? Source directories (src/, app/, lib/, pkg/, …) with real, non-boilerplate code.
  2. Dependency/build manifests? package.json, pnpm-lock.yaml, pom.xml, build.gradle, go.mod, requirements.txt/pyproject.toml, Cargo.toml, *.csproj, Gemfile, etc.
  3. Git history? Zero commits / empty tree vs. a real history.
  4. incu-way already present? A CLAUDE.md that references the incu-way flows, a .ways/ directory, or docs/architecture/.
  5. Branch model? Whether main and develop already exist.

Classify the repo

  • Greenfield — empty or near-empty (no application code yet; maybe just a README or license). The work is to capture intent: product vision and the architecture the team plans to build.
  • Brownfield — an existing application with code and history (assume this is the default). The work is to reverse-engineer understanding: document the architecture and, when warranted, the functional behavior of what already exists.
  • Already initialized — incu-way structure is present. Switch to refresh mode: fill gaps, update stale sections, and (for brownfield) re-run the documentation pass.

Phase 0 exit criteria

Before Gate 1, Claude should be able to state:

  • whether the repo is greenfield, brownfield, or already initialized
  • the detected stack (languages, frameworks, package manager, test runner, build tool) — for brownfield, derived from the manifests and directory layout, not assumed
  • whether main and develop exist (and whether git is initialized at all)
  • which incu-way pieces already exist and which are missing
  • the proposed initialization plan (files to create, branch strategy, documentation depth)

Do not guess the stack or the product purpose. If the manifests are ambiguous or the repo's intent is unclear, note it as an open question for Gate 1 rather than inventing it.

Gate 1 — Confirm mode and scope

Present the detection result and the proposed plan, then confirm with the user.

Show a short summary:

## Initialization plan

- **Mode:** greenfield | brownfield | refresh
- **Detected stack:** {languages / frameworks / package manager / test runner}
- **Branch model:** {exists: main, develop} | {will create: main, develop}
- **Will create / update:**
  - CLAUDE.md (from template, filled with the detected stack)
  - PRD.md (root) — {product vision (greenfield) | reverse-engineered overview (brownfield)}
  - docs/ tree: prds/, bugs/, security/, architecture/
  - .ways/ (work-item state.json lives here later)
- **Documentation depth (brownfield):** {structural only | architecture | architecture + functional}
- **Open questions:** {anything ambiguous about stack or product intent}
  • Ask: "This is the initialization plan. Confirm the mode and the documentation depth before I create anything."
  • Stop. Do not proceed until the user confirms the mode and scope.

Do not treat repo detection alone as approval to scaffold. Gate 1 sets the mode and the documentation depth.

Isolation setup (immediately after Gate 1)

Before writing any file, establish the branch model and ask the user how they want to isolate the init work:

Do you want me to work in a new branch in the current checkout, or create a separate git worktree?

Use the user's answer exactly. Stop after asking; do not create a branch or worktree until the user chooses one.

If the current checkout has uncommitted changes, stop and ask before switching branches or creating a worktree.

Greenfield with no git yet

Use a branch in the current checkout unless the user explicitly asks for a separate worktree after git is initialized. This repo has no commits yet, so main/develop can't exist without one — ask first: "This repo has no commits yet. OK if I create an empty initial commit to establish main?" Only after the user confirms:

git init
git checkout -b main
git commit --allow-empty -m "chore: initial commit"
git checkout -b develop
git checkout -b chore/incu-way-init

Brownfield (git already exists)

If develop is missing, it needs to be created from main (this is a repo-wide decision — confirm at Gate 1 if it wasn't already) and pushed so it exists on the remote. Ask first: "develop doesn't exist yet. OK if I create it from main and push it?" Only after the user confirms:

git checkout main && git checkout -b develop && git push -u origin develop   # only if develop is missing

Option A — Branch in current checkout

# Run from the repo root
git switch develop
git pull --ff-only
git switch -c chore/incu-way-init

All subsequent work — CLAUDE.md, docs, the .ways/ scaffold, commits — happens on this branch in the current checkout. Never write directly to develop or main.

Option B — Separate worktree

Create worktrees only under .worktrees/. Do not ask the user for a path.

# Run from the repo root
git fetch origin develop
mkdir -p .worktrees
git worktree add .worktrees/incu-way-init -b chore/incu-way-init origin/develop

All subsequent work — CLAUDE.md, docs, the .ways/ scaffold, commits — happens exclusively inside this worktree. Never write directly to develop or main.

Phase 2 — Scaffold the incu-way structure

Goal: Create the files and directories the incu-way flows assume, filled with real, detected information (not placeholders, where the repo gives the answer).

2.1 — CLAUDE.md

Copy this skill's bundled claude.template.md (it sits next to this SKILL.md, so it travels with the skill on ways add) to the repo root as CLAUDE.md and fill every [TODO: ...] from what Phase 0 detected:

  • About the project — for brownfield, derive from README, manifests, and code; for greenfield, from the user's vision captured at Gate 1.
  • Stack — languages, frameworks, DB, testing, package manager (detected).
  • Folder structure — the real top-level layout with one-line purposes.
  • Code conventions — infer from existing config (.eslintrc, .editorconfig, tsconfig.json, linters, formatters) for brownfield; from the template defaults for greenfield.
  • Useful commands — read them from the manifest's scripts (package.json scripts, Makefile, pom.xml, etc.). Do not invent commands that don't exist.
  • Keep the mandatory Security, Worktree, Development workflow, and Bug workflow sections exactly as the template defines them.

Anything you cannot derive, leave as an explicit [TODO: ...] and list it at Gate 2 — do not fabricate.

2.2 — PRD.md (root)

The incu-way flows read PRD.md first in their Phase 0. Create it:

  • Greenfield: capture the product vision agreed at Gate 1 — problem, target users, goals, non-goals, the intended high-level architecture.
  • Brownfield: a reverse-engineered product overview — what the application does today, for whom, and its main capabilities, traceable to the code. If incu-way-docs is in scope (see Phase 3) this can be a short pointer to docs/architecture/ and the functional docs, kept deliberately thin.

2.3 — docs/ tree and .ways/

Create the directory structure the flows write into:

docs/
  prds/          # incu-way-development writes here
  bugs/          # incu-way-bugs writes here
  security/      # snyk-remediation writes here
  architecture/  # incu-way-docs writes here (brownfield)
.ways/       # work-item state.json lands here later (add a .gitkeep)

Add .gitkeep to the empty directories so they are committed. Do not create a .ways/state.json here — it belongs to the work-item flows.

.ways/ is the canonical root for way state (ways/v1alpha1). If the project also uses the ways CLI, that tool keeps a cache under .ways/cache/ — see repo hygiene below.

2.4 — Repo hygiene

Always ensure .gitignore has these entries (append whichever are missing; never duplicate a line that's already there):

.claude/
.agents/

.claude/ and .agents/ hold placement output from skills.sh/hooks.sh (and, for .claude/, Claude Code's own local settings) — every bit of it is regenerated from ways.lock / skills-lock.json / steering-lock.json by ways install, so it's exactly as disposable as node_modules: never commit it, and never ask before ignoring it.

Only when relevant to this project:

  • Add .worktrees/ to .gitignore if it isn't already ignored and lives inside the repo.
  • .ways/ gitignore rule. .ways/state.json must stay tracked (resume-by-branch and the board depend on it being committed via incu-way-prepare-pr). If the ways CLI is in use, ignore only its cache — add .ways/cache/ (not the whole .ways/) to .gitignore. Never blanket-ignore .ways/.
  • If the project uses the SonarQube scan referenced by the flows, note where run-sonar.sh lives or that it is still TODO — do not fabricate the script.

Tell the user the scaffold is ready and suggest invoking incu-way-prepare-pr to commit it — do not commit it yourself:

Suggested message: chore(incu-way-init): scaffold CLAUDE.md, docs/ tree and .ways/

Phase 3 — Documentation of the application

Goal: Produce the architectural (and, if warranted, functional) documentation that lets humans and the incu-way flows understand the application.

Brownfield — invoke incu-way-docs

If the documentation depth agreed at Gate 1 is architecture or architecture + functional, hand off to the incu-way-docs skill, instructing it to write into the current init branch/worktree (no separate PR — its output ships with this init PR). incu-way-docs produces, scaled to the app:

  • docs/architecture/ — overview, components, data model, integrations, deployment/ops
  • functional/domain docs and a glossary (when the app is user-facing or domain-heavy)
  • Mermaid diagrams (validated)

If the agreed depth is structural only, skip the deep pass — the CLAUDE.md folder structure + PRD.md overview are enough for now. Note in PRD.md that the deep documentation is deferred and can be generated later with incu-way-docs.

Greenfield — capture intended architecture

There is no code to reverse-engineer. Instead, record intent in docs/architecture/overview.md: the planned architecture, major components, key technology choices and the reasoning (ADR-style). Keep it lightweight; it will evolve as the code is written through incu-way-development.

Gate 2 — Review the initialization

After scaffolding (and documentation, when in scope):

  • Summarize what was created/updated and list every remaining [TODO: ...] in CLAUDE.md and any open questions.
  • Say: "Initialization complete on chore/incu-way-init. CLAUDE.md, PRD.md, the docs/ tree and .ways/ are in place{, plus the architecture/functional docs}. Please review before I open the PR."
  • Stop. Do not open the PR until the user approves.

Phase 4 — Merge via Pull Request

Goal: Land the initialization on develop (and then main) through a reviewed PR. Never merge directly.

PR

After Gate 2 approval, suggest invoking incu-way-prepare-pr to push chore/incu-way-init (including the first push of main/develop if the repo had no remote/history yet) and open the PR. Do not push or run gh pr create yourself.

Suggested PR title/body for incu-way-prepare-pr to use:

Title: chore: initialize incu-way

PR body must include:

  • Mode: greenfield / brownfield / refresh.
  • Created/updated: CLAUDE.md, PRD.md, docs/ tree, .ways/, and (if any) the architecture/functional docs with links.
  • Open items: remaining [TODO: ...] and open questions for the team.

Once opened, share the PR URL with the user. Stop. Do not merge until the user approves.

If the team gates on main too, open the follow-up PR develop → main after the first is merged, same as the development flow.

Quick reference — gates

Gate Trigger What to say Block until
1 Mode + plan detected "This is the initialization plan. Confirm the mode and documentation depth." User confirms
2 Scaffold (+ docs) ready "Initialization complete on chore/incu-way-init. Please review before I open the PR." User approves
3 Reviewed Suggest incu-way-prepare-pr for PR chore/incu-way-init → develop; share URL once user runs it User approves PR

What NOT to do

The phase-level rules above are authoritative; the one point not stated elsewhere:

  • The greenfield/brownfield bootstrap commands are the only exception to the no-auto-commit rule, and even those require asking first — everything else (the scaffold commit, the PR) only happens via incu-way-prepare-pr.