skills

agent
Guvenlik Denetimi
Basarisiz
Health Uyari
  • License — License: MIT
  • Description — Repository has a description
  • Active repo — Last push 0 days ago
  • Low visibility — Only 7 GitHub stars
Code Basarisiz
  • rm -rf — Recursive force deletion command in scripts/test-absent-local-gate-tells.sh
Permissions Gecti
  • Permissions — No dangerous permissions requested

Bu listing icin henuz AI raporu yok.

SUMMARY

Pipeline-first Claude Code skills pack — 30 composable skills for structured feature delivery from shape to ship

README.md

Skill Kit

License: MIT
Skills
CLI

Skill Kit is a pipeline-first Claude Code skills pack for structured feature delivery. 35 composable skills that take a feature from idea to shipped — with explicit handoffs, verified research, and a compounding knowledge loop.

Built around structured research, GitHub-native state, and a compounding knowledge loop in docs/solutions/. Compatible with Claude Code, Cursor, Windsurf, and any agent that can consume SKILL.md files.

Installation

Global install (all skills, Claude Code):

npx skills@latest add chrislacey89/skills --skill '*' --agent claude-code --global -y

Requirements

Most skills need only an agent that reads SKILL.md files. The pipeline's GitHub-native state and the sidecar resolver set these version floors:

Requirement Needed for Notes
gh ≥ 2.94.0 issue dependencies and sub-issue relationships in /prd-to-issues, /qa, /execute, /help, /setup-ralph-loop Released 2026-06-10. Earlier versions lack gh issue edit --add-blocked-by and the blockedBy / blocking --json fields. Check with gh --version.
git ≥ 2.31 the sidecar resolver (sidecar.sh) bundled into /compound and /research Earlier versions lack git rev-parse --path-format=absolute, which the resolver uses to find the main checkout from a linked worktree. Check with git --version.

On GitHub Enterprise Server, issue types and sub-issues need GHES 3.17+, and issue dependencies need GHES 3.19+. Where dependencies are unavailable, the filing skills degrade to the prose ## Blocked by section and say so rather than failing.

Keeping skills updated

The skills CLI tracks content hashes. When this repo updates, your installed copies know about it:

npx skills@latest check    # see what changed upstream
npx skills@latest update   # pull latest versions

What Skill Kit optimizes for

  • GitHub-native state — PRDs, slices, QA bugs, and lineage live in GitHub issues and PRs, not a sprawling local planning filesystem. On a guest repo, see docs/state-backend.md.
  • Explicit handoffs — skills declare what they expect, what they produce, and what comes next.
  • Verified research before PRD writing — /research is a first-class step, not an optional extra.
  • Compounded knowledge — shipped lessons feed future work through docs/solutions/.

Canonical pipeline

/shape → /research → /write-a-prd → /prd-to-issues → /execute → QA → /pre-merge → /compound (in-PR) → /pre-merge (re-run: re-stamp) → /closeout (merge + teardown) → cleanup

The pipeline is the default path, not a prison. Skills can backtrack when assumptions fail or branch to helper and side-route skills when the work demands it. For blank-project or major-tranche work that is too large for a single PRD, /shape can branch to /create-milestone, which creates a GitHub milestone plus feature issues that mature from roadmap bet to research-ready to prd before re-entering the normal pipeline at /research. Ralph is the AFK execution mode/persona for the /execute stage, not a separate pipeline step — it stops on repeated failure or repeated non-progress, whichever trips first.

Skills

Planning & Design

Skill Description
shape Structured requirements discovery — shared understanding before research
create-milestone Turn a shaped app-sized or tranche-sized idea into a GitHub milestone with sequenced feature bets
research Mandatory pre-PRD research with auto-calibrated depth
write-a-prd PRD creation via interview, filed as GitHub issue
prd-to-issues Break PRD into vertical slices with boundary maps
design-an-interface Generate multiple radically different interface designs
api-design-review Focused contract review for higher-risk API design decisions
prototype Throwaway code that answers a question — LOGIC (state model), UI (layout variants), or FEASIBILITY (spike-solution discharge for Uncertain research assumptions)
mermaid Generate Mermaid diagram source (flowcharts, sequence, ER, state, C4, Gantt, and 18 more) for embedding in PRDs, research notes, and issues

Development

Skill Description
execute Execute a unit of work end-to-end with verification
tdd Test-driven development with red-green-refactor loop (invoked from /execute)
fix-findings Fix the review findings you chose (from /pre-merge, or teammate threads from /respond-to-review), with a fresh sub-agent writing each fix and a second one trying to break it — run from the /pre-merge menu or typed, never selects findings, never stamps
triage-issue Deep bug diagnosis + root cause + TDD fix plan (invoked from /qa per issue)
improve-codebase-architecture Surface deepening opportunities for shallow modules
request-refactor-plan Plan refactors with tiny commits
ts-audit Audit TypeScript code against 9 bundled TypeScript library references
lfg Take an ask to a proof of concept: short shaping conversation, then unattended build of tracer-bullet slices to a stamped PR with a recap and gap issues. User-invoked only, never merges, never stamps. Send /lfg as its own message — typed as the first message of a fresh Conductor workspace it is not parsed as a command, and the run proceeds without the skill
hackathon Plan, build, and demo one project inside a judged hackathon's build window: an event profile that marks every unstated rule not stated, candidates scored dimension by dimension with a blind second scorer, an over-full scope ladder re-planned from measured rung pace at fixed clock checkpoints, a production-grade build with a visible factory, and a table demo with judge Q&A drilled from a decision log. User-invoked only. Send /hackathon as its own message

Tooling & Setup

Skill Description
init-pipeline Scaffold pipeline enforcement — Claude Code hooks, git guardrails, pre-commit setup (auto-invoked by /execute)
setup-pre-commit Commit-time quality gates via Lefthook; see its SKILL.md
setup-ralph-loop Generate ralph-once.sh and a bounded ralph.sh for HITL-to-AFK /execute runs (auto-invoked by /execute for multi-slice work)
git-guardrails-claude-code Block dangerous git commands

Knowledge & QA

Skill Description
qa Single entry point for bug conversations; files lightweight issues and delegates per-issue to /triage-issue for deep diagnosis
pre-merge Author-mode: create the PR and run an architectural review before merge. Reviewer-mode (--pr <n>): review someone else's PR and draft comment text. Loop-mode (--loop, or auto-entered on an AFK /execute handoff): records every finding in a durable ledger on the PR with the evidence for or against it, so findings survive the session — makes no commits and no decisions; the operator settles each row
walk-commits Interactive commit-by-commit comprehension walkthrough before merge — intent, riskiest line, deliberate oddities, what's absent by design, per-commit sign-off (optional, recommended by /pre-merge; can render per-commit callouts via the shared visual-rendering core)
visual-recap Render a finished diff/PR/branch as a self-contained interactive HTML artifact in one of two modes — a scrolling recap (file-tree + change flags, annotated split diffs with line-anchored callouts, contract cards, UI wireframes, before/after columns, a per-unit series when one root cause repeats across four or more near-identical sites) or a walkthrough deck (premise → changes → mechanism → aftermath, one idea per screen); CSS-spine diagrams for trivial flows, Mermaid-via-CDN for multi-stage graphs, copy-text feedback loop (optional, never auto-invoked, transient artifact)
respond-to-review Respond to a teammate's review of your PR — a fresh sub-agent checks each comment against the code, you decide fix / push back / postpone / answer, fixes go to /fix-findings, pushback gets a reply draft you post, and a record stays in the sidecar
closeout Merge the reviewed PR, tear down the worktree, prune the branch, and return to a clean base
compound Capture lessons learned into docs/solutions/ — onto the open PR, handed back to /pre-merge so the entry is reviewed and re-stamped before /closeout merges (default), or post-merge as the fallback
excavate Mine a legacy repo's last 12 months of git history (and PR threads via gh) into ranked hotspots, cited recovered decisions, and a who-to-ask map, then offer a PR of docs/solutions/ entries the pipeline already reads. User-invoked only (disable-model-invocation: true); never merges
improve-pipeline File a Skill Kit improvement proposal as a GitHub issue when the lesson is about the pipeline rather than the project
ubiquitous-language DDD glossary. User-invoked only (disable-model-invocation: true)

Orientation & Recovery

Skill Description
help Read repo state and recommend the next pipeline step
correct-course Diagnose stale artifacts and walk the cleanup before backtracking
handoff Compact the current session into a transient handoff doc when no pipeline artifact fits
re-pitch Re-state an explanation that did not land — diagnose the confusion, rewrite in controlled technical English, gloss every domain term on first use

Repo guide

License

MIT

Yorumlar (0)

Sonuc bulunamadi