easyhaloo/afk

Create, review, or revise an AI-Friendly Product Requirements Document in Markdown. Use when a user wants a PRD that an AI can parse, question, decompose, implement, and verify. The skill defines fixed Markdown sections, atomic requirements, Mermaid diagrams, stable IDs, traceability, and a self-check gate. Do not use it to create provider records or execute backlog items.

Cross-branch code migration — when you need to cherry-pick code from one branch to another with large differences. Analyzes commit content, classifies core vs incidental changes, assesses conflict risk, and guides step-by-step migration with rollback support. Pure Git only.

Plans and adds focused regression tests for the business flow affected by a existing code that lacks adequate unit, integration, or E2E tests. Adds tests incrementally without changing production behavior. Uses parallel read-only exploration and a scoped TEST-PLAN.md; not for feature development, bug fixing, or repository-wide coverage projects.

Use when a specific, reproducible failure is provided (command, curl, or sequence) and root-cause diagnosis + fix verification is needed. Executes in current session with human confirmation at each step.

Use when user requests a specific coding task to execute in the current session (e.g. "帮我完成 X", "add login", "fix null pointer"). HITL execution within the session. Optional escalation to background execution upon completion.

Requirements interview skill — use when a feature/epic is ambiguous and the team needs shared, falsifiable understanding before building. Runs multi-round HITL interviews via AskQuestion. Produces a structured CONTEXT.md with Audience, Success Criteria, Non-goals, Constraints, and Open Questions. The interview has a non-negotiable floor: at minimum, one Audience, one Success Criterion, and one explicit Non-goal must be stated — no build proceeds below this floor. Output goes to /tmp/, never to the repo working tree.

Gap-fill interview skill — use when bounded contexts, architecture docs, a prior draft, or code audit results already exist and need verification, correction, or expansion through targeted questioning. Starts from what is already known, digs into holes. May read code to verify context against the actual codebase. Produces a revised CONTEXT.md via the same Step 4 gate (Approve / Revise / Drill / Add Open Question). Output goes to /tmp/, never to the repo working tree.

Use when the current session should snapshot its state so the next session resumes immediately without re-explaining context. Save mode writes to /tmp/; resume mode reads and continues.

Use when a backlog item with machine-checkable Acceptance Criteria is ready for autonomous implementation before independent QA.

Use when user asks about LLM application design, architecture, skill-driven systems, context engineering, HITL, observability, or evolvability.

Generate product prototypes or technical spikes. - product: HTML pages from vague requirements → iterate until confirmed - spike: prove technical feasibility → draft provider change + findings

Use when a backlog item is in verification and needs independent Acceptance Criteria checks before the QA runner publishes its PR/MR.

Understand how an existing system works, or evaluate feasibility of an approach, before committing to a plan. Outputs findings doc or spike summary.

Use when multiple backlog items need dependency-aware autonomous execution. Plans runnable waves and starts the provider-backed loop.

Create a new SKILL.md, diagnose an existing one for quality issues, or refactor to align with SKILL-GUIDE standards. Trigger: user asks to create, audit, or improve a skill.

Decompose approved requirements into a provider-neutral backlog manifest with machine-checkable acceptance criteria, dependencies, parent links, execution modes, and business tags. This skill drafts and validates the handoff; an external backlog system creates the records.

Synthesize aligned requirements into a provider-neutral PRD for publication and later external backlog decomposition. Produces problem, users, bounded contexts, user stories, decisions, risks, and non-goals without creating provider records or invoking AFK execution commands.

Multi-step API testing with Playwright. Generates executable test files from business scenarios with auth reuse. Pick the auth mode by where the test runs: local dev uses a long-running Chrome on 9222 + CDP attach (default); CI uses storage-state from API login. Triggers: "test order flow", "verify webhook", "trigger async job, poll for completion", "API then browser verify", "reuse browser login", "skip UI login", "browser auth session".

Globally install Jev checkpoints for Claude Code, Codex, and OpenCode. Recommend relevant Skills, monitor observable execution evidence, and reduce safe tool-output noise without replacing host permissions.

Convert Markdown documents (with Mermaid diagrams, tables, Chinese text) into polished A4 PDFs. Handles ```mermaid code blocks by rendering them to PNG automatically, and applies professional typography suitable for Chinese + English mixed content. Use this skill when: - User asks to "convert this markdown to PDF" - User wants to "export" a .md file as PDF - User needs to print, share, or present a Markdown document - User needs diagrams (Mermaid) rendered inside a PDF - User says "生成 PDF", "转 PDF", "导出 PDF"

Use when coding agent exhibits reasoning path failures during multi-turn coding sessions: error accumulation across iterations, repeated edits on the same location without convergence, excessive token consumption without measurable progress, or semantic regression where code quality degrades across edits. Detection is session-context-based (no hooks required); the agent applies correction prompts proactively within the current conversation. Also use when asked to "apply reasoning guard", "check my reasoning path", "stop and rethink", or "am I going in circles". Do NOT use when the agent is producing correct output steadily.

Automated hook-based reasoning path guard for coding agents. Installs PostToolUse/PreToolUse/SessionEnd hooks into Claude Code to detect error accumulation and intercept before damage compounds. Use when the reasoning-watchdog system is already installed (npm run install completed) and you need to check status, tune thresholds, or troubleshoot. Do NOT use for session-context-based reasoning monitoring — this skill is hook-based, not conversation-based.

Assess and govern software complexity from method to repository scope. Use when measuring cyclomatic or cognitive complexity, code smells, coupling, module or service boundaries, duplication, change hotspots, or technical-debt priority.

Manage a local Obsidian Markdown wiki: ingest reusable knowledge from the current conversation and verified task evidence, lint pages, or search and read existing knowledge. Use for Obsidian vault and local wiki requests; LLM Wiki desktop API and cloud knowledge bases use their own tools.