ワンクリックで
claude-harness
claude-harness には poz110 から収集した 20 個の skills があり、リポジトリ単位の職業カバレッジとサイト内 skill 詳細ページを表示します。
このリポジトリの skills
全流程自動模式:從當前狀態一路推進到 DONE,無需人為干預確認。自動推進 MANUAL 節點、派發 Agent、處理失敗回滾,直到項目完成。支持傳入需求描述,自動注入後續流程。
快速修復模式:徹底分析問題後直接寫修復代碼。不走 pipeline,不生成文檔。適用於緊急 bug 修復或小範圍改動。
统一的 Jira 问题处理中心。提供:ensureConnected() 确保连接可用(MCP/curl 双通道 + 多源凭证 + 缓存)、getIssue() 获取完整 ticket 上下文(含图片分析)、writeBack() 回写结果到 Jira。被 autopilot/hotfix/feature 调用,所有 Jira 相关逻辑集中在此。
增量功能模式:等同於 /autopilot feature,跳過 Arch/Design 階段,適合在現有項目上添加新功能。
Universal environment checker. Detects missing tools, services, and configurations BEFORE starting any work phase. Always asks user before installing anything. Used by Designer (Stitch), FE, and BE agents. Modules: A = Stitch MCP (Designer), B = Backend env (BE), C = Frontend env (FE).
Implements backend APIs using Bun, Hono, Drizzle ORM, tRPC v11, better-auth. Load this skill when BE agent needs implementation details.
Implements frontend features using Next.js 15, React 19, shadcn/ui, Tailwind v4, tRPC/TanStack Query. Load this skill when FE agent needs implementation details.
Sets up GitHub Actions CI/CD, Dockerfile, deployment configuration, and monitoring. Used by DevOps agent after SECURITY_REVIEW passes.
Full ADR generation with mandatory ASCII diagrams (data flow, state machine, sequence, error paths), tech debt radar, and security baseline. Used by Architect agent after PRD approval.
Architecture compliance code review. Used by Reviewer agent after IMPLEMENTATION. Two-stage sequential review: spec compliance first, then code quality. Each stage must pass independently. Produces PASS/FAIL verdict.
Executes the Designer workflow: AI Slop detection → competitive research → design system proposal → design spec generation. Produces DESIGN.md and docs/design-spec.md with 80-item audit score.
Executes the full PM workflow: office-hours interrogation → scope-mode selection → PRD generation. Must not skip office-hours step.
Generates docs/interaction-spec.md through a TWO-PHASE process: Phase A — Designer extracts every interactive element from design screens and presents an Intent Confirmation Checklist to the user (DESIGN_PHASE MANUAL node). Phase B — after user confirms intent for each element, Designer locks the confirmed decisions into interaction-spec.md, which becomes the authoritative behavioral contract for FE, QA, and Reviewer. The spec is not a template to fill in — it is a record of confirmed human intent. Invoke at DESIGN_PHASE MANUAL after CEO_REVIEW completes.
啟動工作流可視化監控面板,實時顯示狀態、耗時、產物文檔。HTTP + WebSocket 服務,支持多客戶端同時連接。
OWASP Top 10 (2021) security audit. Read-only scan using Grep and Bash. Cross-references docs/security-baseline.md for project-specific constraints.
Creates test plan, writes Vitest unit tests, Playwright E2E tests, and k6 performance tests. Used by QA agent after CODE_REVIEW passes.
Generates real UI design screens using Google Stitch MCP. Derives complete page list from PRD + design-spec, builds per-page prompts with full feature context, calls Stitch MCP, verifies output completeness, and hands off to FE. Never silently degrades or skips pages. Always confirms page inventory upfront.
Structured debugging methodology for implementation failures. Use when a build fails, tests fail, or behavior is unexpected. Replaces ad-hoc guessing with hypothesis-driven root cause analysis.
Generates and maintains the requirement traceability matrix (docs/traceability-matrix.md). Maps every PRD Must/Should feature (F###) to its expected implementation location, Gherkin scenarios, test IDs, and completion status. Required output of PRD_REVIEW phase. Consumed by FE, BE, Reviewer, and QA. Invoke when: Architect completes ADR, FE/BE update implementation status, Reviewer verifies coverage, QA links test IDs.
Methodology for creating new workflow skills. Enforces TDD for process documentation: write failing scenario first, then write the skill. Use when adding any new SKILL.md to the workflow.