@weloin/claude-skills

Weloin's Claude Code skills library — install with one command, updates via symlink. macOS · Linux · Windows.

weloin:project-setup

A phase-machine for agent-driven development. Instead of ad-hoc prompting, it runs your project through structured phases — from first interview to shipped code — with every decision recorded and resumable.

What you get

• An interview that produces your brief, requirements, architecture, and ADRs.
• A custom agent roster with model tags matched to each role.
• Phased plans executed at your chosen AUTO / GUIDED / MANUAL autonomy.
• A docs/project/STATE.md single source of truth — close a session mid-phase, resume days later without re-explaining anything.
• Works on fresh directories or aligns an existing codebase to the same structure.

Why it holds up

• Durable state. Agent sessions lose context; this makes project state survive across them.
• Maturity axis (prototype → mvp → production) sets the quality bar independent of scale — prototypes skip ceremony but log every shortcut to a debt ledger; promote later upgrades by re-asking skipped questions and building a hardening plan from that ledger.
• Evidence-based gates. Opt-in quality gates mean work isn't "done" until it passes review with proof — not when the agent says so.
• Strategy that fits. A 15-strategy catalog with a project-kind recommendation matrix, not one-size-fits-all.
• Configurable Git workflow. Commit strategy (style · detail · AI-signature) and worktree policy (per-phase / per-feature / ask / none) set once, applied automatically.
• Never frozen. Every config field has a subcommand to re-set it later; changes apply going forward, never restructuring existing work.

Full subcommand reference & opt-in features → docs/project-setup.md.

Quick install

npm install -g @weloin/claude-skills   # then:
weloin-skills                          # pick skills interactively

Full install, update, and CLI reference → docs/installation.md.

What execution looks like (per feature / fix)

Once set up, every substantial change runs the same loop — 🧑 = you decide, 🤖 = Claude acts. Config (worktrees, autonomy, gitflow, commit_strategy) steers the branches without re-asking.

flowchart TD
    Start([🧑 Request: feature or fix]) --> Trivial{🤖 Substantial change?}
    Trivial -->|no, one-liner| CurCheckout[🤖 Work on current checkout]
    Trivial -->|yes| Policy{🤖 worktrees policy?}

    Policy -->|per-feature| MkWT[🤖 Auto-create worktree + branch]
    Policy -->|per-phase| Resolved[🤖 Already isolated at macro-plan]
    Policy -->|none| CurCheckout
    Policy -->|ask| AskIso[🧑 Choose: current branch / new branch / worktree]
    AskIso --> MkWT

    MkWT --> Plan
    Resolved --> Plan
    CurCheckout --> Plan

    Plan[🤖 Write plan: goal, files, tasks, DoD] --> MacroAsk[🧑 Macro-plan ask: gitflow branch? autonomy mode?]
    MacroAsk --> Gate1{🧑 Approve plan?}
    Gate1 -->|revise| Plan
    Gate1 -->|approve| Loop

    subgraph Loop [Task loop — per task]
        direction TB
        Dispatch[🤖 Builder subagent: task + spec + rules] --> Verify[🤖 Verify vs maturity DoD]
        Verify --> Review[🤖 Reviewer subagent]
        Review --> Gates{🤖 Quality gates at group boundary?}
        Gates -->|FAIL| Fix[🤖 Fix loop] --> Dispatch
        Gates -->|PASS / CONCERNS| Commit[🤖 Commit per commit_strategy + progress + STATE.md, same commit]
    end

    Commit --> Mode{🧑🤖 Autonomy mode?}
    Mode -->|AUTO| More
    Mode -->|GUIDED, group done| PauseG[🧑 Review report → go]
    Mode -->|MANUAL| PauseM[🧑 Approve next task]
    PauseG --> More
    PauseM --> More

    More{🤖 More tasks?} -->|yes| Dispatch
    More -->|no| Merge[🤖 Merge per gitflow + clean up worktree]
    Merge --> Done([🤖 Update STATE.md next: → resume-ready])

Deeper walk-through → docs/execution-flow.md.

Docs

• Installation & CLI — install, update, uninstall, command/flag reference
• project-setup reference — subcommands, opt-in features
• Execution flow — per-feature loop, human ↔ Claude gates
• Development — how installs work, adding a skill, troubleshooting
• Changelogs

License

MIT © Weloin