Browse reusable Agent Skills, each with a clear purpose and practical guidance.
This skill should be used when needing to locate files within the Claude Code plugins cache directory (~/.claude/plugins/cache). Triggers include finding tool scripts, skill files, or any plugin resource when the hardcoded path is unknown or varies by plugin version. Use when slash commands or orchestrators need to dynamically resolve plugin file paths.
蒸馏一个你身边的人。通过聊天记录、朋友圈、描述等素材，生成 ta 的人格档案，让 ta 以自己的方式和你对话。
从过去 N 天的对话历史中提炼经验教训——分析每次 agent 交互中的异常、失败、成功模式， 提出 CLAUDE.md/AGENTS.md 更新建议、可凝聚为 skill 的新方向、以及现有 skill 的改进点。 当用户说"总结历史对话"、"learn from history"、"回顾最近的对话"、"从历史中学习"、 "分析最近的 agent 表现"、"对话历史分析"、"最近有什么可以改进的"、 "提炼一下这几天的经验"、"历史经验总结"时立即触发。 也适用于用户想知道"最近 agent 出了哪些问题"、"有什么可以优化的"、 "哪些操作可以自动化"等经验提炼场景。 只要用户想从对话历史中系统性地提取可行动的改进方向，就应使用此 skill。
Per-node MEMORY slices (Leg A — self/history) — FIND a node's standing lessons + cross-run RECURRENCE before the optimizer edits it, and MAINTAIN the memory.md set so it stays honest, bounded, and never rots. The Leg-A twin of `okf-slices` (Leg B — world/code): a memory lesson `[[links]]` the OKF code slice it concerns, so its freshness gate rides that slice's `--check`. TRIGGER on either intent: (FIND) the out-of-band triage/fixer needs "has this node failed THIS way before, how often, and what did we learn?" — the LAPSE-vs-SKILL signal — before bucketing or editing a node; or (MAINTAIN) someone runs the MEMORIZE step, asks when/how memory.md is updated, what a lesson's scope is, whether a lesson is stale, or WHERE a lesson should live (per-node vs system). Works on any repo with template-scoped `memory.md` (`.piflow/<wf>/template/{memory.md, nodes/<id>/memory.md}`; config in `.agents/okf/okf.config.json`). Memory is OPTIMIZER-FACING reference, NEVER injected into a worker node's runtime prompt — this skill is how the out-of-band optimizer reads and writes it.
Generate new workflow nodes with implementation, tests, and registry registration
Canonical reusable orchestration procedures for the ClosedLoop planning/implementation orchestrators — PLAN_VALIDATION_SEQUENCE and AWAITING_USER_SEQUENCE. Single source of truth shared by prompts/prompt.md (full loop), prompts/plan-prompt.md (/code:create-plan) and prompts/execute-prompt.md (/code:execute-implementation), which each load standalone and cannot reference one another at runtime. Triggers: - An orchestrator phase needs full plan validation (structural + semantic) — run PLAN_VALIDATION_SEQUENCE. - An orchestrator reaches a hard stop that requires user action before continuing — run AWAITING_USER_SEQUENCE.
Pi Flow — author, enhance, and run structured filesystem-coordinated workflows (a DAG of producer/verify nodes) as a fleet of efficient pi agents driven by non-Claude coding-plan models, with Claude Code as the single console. This is the plugin OVERVIEW/router; the work lives in three skills — piflow-init (create), piflow-enhance (improve), piflow-start (run). Source of truth is a structured workflow TEMPLATE; the `@piflow/core` SDK loads it and runs one `pi` per node.
Pi Flow · ENHANCE — improve an existing, running pi-flow workflow the disciplined way: turn a spotted flaw, a recurring finding, or human feedback into ONE atomic, generalizing change via capture→route→edit→verify→ approve→commit, never a one-off hack. Owns the criteria fixture (the per-node quality bar) and Companion-Mode judging. Use to "improve a wave/node", "fix a recurring failure in the pipeline", "the generated output is wrong — fix the system not the case", "edit the skill vs edit the chain", "score / triage / fix a finished run", "run the optimize loop", "auto-improve a node from its trace". Also owns the METHOD-LIBRARY sync (Leg C, ~/Desktop/best-designs-for-agentic-system): consult the matching card before any edit, write the outcome back to its Applications after any land. To CREATE a workflow use piflow-init; to RUN one use piflow-start. The canonical capture→route→edit method is the hermes-skill-system skill; this COMPOSES it and pins the piflow-specific precedence + the autonomous SCORE→TRIAGE→FIX→GATE→LAND loop.
Pi Flow · OVERLORD — the CONTROL-PLANE agent contract: the seat between the human and the pi fleet that OBSERVES the canonical telemetry stream, DECIDES, and ACTS through piflowctl (run / optimize / fix), making the high-order calls a deterministic controller can't — continue · abort · rerun-with-a-steer · nudge · escalate · land. LOAD THIS when you are SUPERVISING a live run, an optimize pass, or a fix loop and must judge a node's behaviour and decide its fate; when you are "spawning + testing" a fixer/agent and deciding whether to shut it down, re-run, or escalate; or when "act as the control plane", "be the overlord / the governor", "supervise / babysit this run", "should I kill / rerun / escalate this", "insert a control plane", or "k8s-style control of the flow" come up. The overlord has TWO MODES — a programmatic controller and an agent (you) — fed by the SAME stream; this is the AGENT-mode contract. It DELEGATES deterministic termination (timeouts, retries, the run-count ceiling, token/edit budgets) to the shipped workflow-management plane and adds JUDGMENT on top; it intervenes AT SEAMS for live producer runs and may abort mid-run ONLY off the critical path (a candidate / control node). It is an event-WOKEN SENTINEL with NO event loop — a declarative wake-policy over the one stream decides when a fold event is owed human-grade judgment, then it wakes, adjudicates a one-shot digest, acts at a seam, and sleeps. piflow-start is its actuator for running & monitoring; piflow-inspect is its telemetry/instrument source (status · telemetry · trace · logs); piflow-triage + piflow-fixer are the two agents it spawns and supervises over an optimize round (name the defect, then solve it) — this skill is the decider ABOVE all four. THIS IS THE DEFAULT SEAT WHILE ANY RUN, OPTIMIZE PASS, OR ROUND IS LIVE: load it whenever you are monitoring, supervising, or babysitting a pi-flow run or fix loop — not only on an explicit "act as control plane" ask. LOAD THIS ALSO when you are running triage/fix and must decide continue/abort/rerun/escalate/land; when "doing a round" or "monitoring an optimize pass" comes up; or when "supervise / babysit this run", "monitor this optimize round", "should I kill / rerun / escalate this", or "insert a control plane" come up.