Agent Skills

---

name: evaluate

description: Evaluate an AI workflow against representative cases and its permitted actions. Use for model, retrieval or agent evaluation; tests do not grant release authority.

---

name: fde

description: Keeps the engagement record for client work. Use when they name a client or stakeholder. Use when they debrief a meeting or paste notes. Use when they ask what was agreed. Use when they run a POC, change the client's codebase, prove it on their staging, go live, or need evals before a model acts. Use when they prep a readout, when trust shifts, or they say @fde. Route and run the local fde CLI (or npx --yes fdeops). Never ask them to type commands. Not for ordinary code edits in an unbound repo.

---

name: feedback

description: Assess a field lesson for reuse or product feedback without exposing customer context. Use for recurring deployment lessons; distinguish a hypothesis from a validated pattern.

---

name: handoff

description: Transfer operation of a customer deployment with ownership, evidence and a tested support path. Use for handoff or an engineer rotation, not merely code delivery.

---

name: ingest

description: Fetch requested source material and prepare sourced engagement updates. Use to catch up from external notes or messages; applying updates requires a bound record and confirmation.

---

name: integrate

description: Build or change a customer-system integration with explicit data mapping, permissions, retries and reconciliation. Use for connectors, imports, write-back and upstream APIs.

---

name: options

description: Compare feasible approaches to a customer problem and recommend a path with costs, constraints and evidence. Use for an architecture or delivery decision, not implementation.

---

name: plan

description: Sequence an understood outcome into verifiable delivery slices with dependencies, ownership and acceptance checks. Use for delivery planning, estimation or migration strategy.

---

name: poc

description: Run a bounded customer proof of concept to test a consequential uncertainty. Use for a spike or pilot with a question and decision deadline, not a full rollout.

---

name: prioritize

description: Choose up to three immediate priorities from competing initiatives. Use when everything is urgent and the customer needs a defensible order with explicit deferrals.

---

name: qa

description: Exercise the delivered customer journey using real runtime or browser evidence. Use for functional acceptance testing after implementation, including failure paths.

---

name: readout

description: Prepare a sponsor update separating promised outcomes, measured results and customer acceptance. Use for progress readouts or defending a delivery claim.

---

name: red-team

description: Stress-test a plan, brief or delivery claim against evidence and plausible failure modes. Use when the user asks for a red team or preparation for a consequential decision.

---

name: rescue

description: Triage an outage, loss of stakeholder trust or a failing engagement direction. Use for urgent recovery; distinguish diagnosis from authorized production mitigation.

---

name: review

description: Review a proposed customer code change against its intended outcome and operational risks. Use for a diff or PR review; report evidence and actionable findings.

---

name: rollback

description: Prepare and rehearse a recovery path for an intended release. Use when rollback is assumed, untested or previously failed; follow the environment authority for any drill.

---

name: runbook

description: Write an operating runbook from the delivered system and verified procedures. Use when the customer team or a successor needs to operate without the original engineer.

---

name: scope

description: Assess a new customer request against agreed scope, trade-offs and ownership. Use when an engagement expands or a custom feature needs a commitment decision.

---

name: score-use-cases

description: Compare competing customer use cases by value, feasibility and evidence. Use when several problems compete for delivery capacity.

---

name: llm-pipeline-analysis

description: Inspect LLM torch profiler traces at forward-pass, layer, and kernel level. Use when you need layer timings, anchor-kernel boundaries, representative kernel flows, or Perfetto time ranges.

---

name: llm-serving-auto-benchmark

description: Framework-independent LLM serving benchmark skill for comparing SGLang, vLLM, TensorRT-LLM, TokenSpeed, or another serving framework. Use when a user wants to find the best deployment command for one model across multiple serving frameworks under the same workload, GPU budget, and latency SLA.

---

name: llm-serving-capacity-planner

description: Parse SGLang/vLLM startup logs to explain GPU memory use and request capacity. Use for KV cache budget, mem-fraction-static comparisons, OOM triage, and max-concurrency estimates.

---

name: llm-torch-profiler-analysis

description: Analyze existing Torch Profiler Chrome traces or capture from SGLang, vLLM, TensorRT-LLM and TokenSpeed servers. Return kernel attribution, overlap opportunities and fusion patterns; use for model dispatch, CUDA Graph gaps, stream contention or exposed kernel tails. Includes existing SGLang Omni traces.

---

name: luban

description: 鲁班（Luban）——Skill打磨工坊。把一个"能用的Skill"打磨成"能被理解、能被安装、能被传播、能被验证、能持续进化"的公共Skill资产。 方法论是工匠式的五个动作：验料（先挑战这个Skill的前提是否成立，不值得雕的料直说）、访行（联网寻找同类Skill，看清自己在生态里站什么位置）、过尺（结构、实测、活体三把尺一起量——活体指拉真实运行产物对账，绿色的CI会撒谎）、慢刨（冻结原版做基线，改动必须通过验证门才保留，否则回刀；验证手段尽量沉淀为仓库里的工具和规矩）、回炉（发布不是终点，留对标观察清单，下一轮从真实反馈进）。 当用户想要升级、优化、打磨、产品化、发布自己开发的Skill时使用。最终产出一份结构化的《Skill打磨报告》、可直接替换的改写片段，以及一张可截图传播的"出师证书"结果卡。 触发词包括但不限于：让鲁班看看这个skill、班门打磨、打磨我的skill、升级我的skill、优化这个skill、skill体检、skill审计、产品化我的skill、这个skill怎么发布、对标一下同类skill、为什么我的skill没人装、帮我把skill发到GitHub/ClawHub、改进SKILL.md。 即使用户只是丢来一个Skill目录、GitHub仓库链接或一段SKILL.md说"帮我看看怎么改"，只要上下文是想让Skill变得更好用、更可传播，都应该触发。 不要用于从零创建一个新Skill（用skill-creator）、不要用于普通的代码review（用code-review）、不要用于改写一段和Skill资产无关的普通提示语。