Build_Great_Loop

Author a NEW, ready-to-run prompt or harness for a long-running iterative agent after its goal and load-bearing decisions are settled. Use for iterate-until-passing, multi-session, self-verifying, research, coding, or bulk processing loops that need state, stop conditions, and recovery. If intent is still unsettled, use loop-spec first. If a harness already exists and the user wants it improved or diagnosed, use loop-review first. Use loop-eval for test design, loop-ops for scheduling/unattended operation, and loop-retro after a real run. The deliverable is the smallest sufficient paste-ready harness, not the task itself. Trigger only when the user explicitly wants a reusable loop prompt/harness; direct requests to fix, run, or iterate on the task itself are ordinary execution. Requests that include actual scheduling/operation are led by loop-ops; when called by loop-ops, return only the per-run prompt.

Design success criteria and a small eval set for an agentic loop — so the loop knows when it's done, and so YOU can measure whether the loop is any good. Use this whenever the user asks how to know if their agent/loop is working, how to define or sharpen success criteria, how to write evals/tests/graders for an agent, how to measure agent reliability, or wants pass/fail checks for a long-running task. Also use as the verification step when building a loop with loop-engineering. Covers verifiable criteria, code-based vs model-based graders, reliability metrics (pass@k vs pass^k), and balanced positive/negative cases. For a new unscoped loop, loop-spec owns success-definition intake first. Mining failures from specified run artifacts is led by loop-retro. Whole-harness audits are led by loop-review. This skill owns a standalone eval artifact once scope is settled. Routing tests for this suite are owned by loop-review.

Operate an existing or intent-settled long-running, recurring, or unattended agent loop SAFELY — the scheduling, cost, safety, and rollout layer around a loop. Use this WHENEVER the user wants to RUN/SCHEDULE/OPERATE an agent loop over time rather than author its prompt: "run this agent every day / on every PR / nightly / on a cron", "set up an autonomous agent that keeps running", "babysit my PRs", "auto-triage issues", "sweep CI failures", "draft release notes automatically", "how do I run my loop unattended without it doing damage", or any question about an agent loop's cadence, autonomy level, token/cost budget, safety gates, denylist, human escalation, multi-loop coordination, kill switch, or operational readiness. If the per-run goal, success, safety, or budget is unsettled, loop-spec owns the request first. For a compound build-and-schedule request, loop-ops is lead and calls loop-engineering once for the per-run prompt. Produces an operating plan (pattern + cadence + autonomy level + budget + safety g

Post-run retrospective for an agent loop — close the improvement flywheel. Reads the ACTUAL run artifacts (structured event log / progress view / state files / PLAN history / loop-docs) and produces: an evidence-cited diagnosis, concrete harness revision proposals, gotcha eval cases from real failures, and SPEC/STANDARDS revision proposals for human ratification. Trigger AFTER a loop has run: "loop 跑完了帮我复盘", "看看运行记录哪里出问题", "这个循环为什么 thrash / 超预算了", "分析一下 run log", "post-mortem my loop", "did my loop actually do a good job". Not for auditing a loop prompt BEFORE it runs (loop-review), operating/scheduling a loop (loop-ops), or building one (loop-spec / loop-engineering).

Audit and harden an EXISTING agentic-loop prompt or harness against loop-engineering best practices. Use this whenever the user has an agent/loop iterative/autonomous prompt and wants it reviewed, critiqued, scored, or made more reliable — or when they describe loop misbehavior and want to know why: "my agent runs forever / never stops", "it quits too early", "it says it's done but it isn't", "it keeps redoing the same thing", "it forgets progress after a restart", "it edits the tests to pass", "review/improve my iterative agent prompt", "why is my loop flaky". It produces a scored report — critical gaps first, each with a concrete fix — and can rewrite the prompt to close them. Exclude one-shot chatbot/system prompts. If symptoms belong to a completed run or run artifacts, loop-retro diagnoses first. Whole-harness audits are led here even when criteria are the suspected defect. This skill also owns maintenance audits of this six-skill router.

Clarify and scope a NEW agent loop before its harness is authored. Use when the user wants to create an autonomous/repeating agent, asks to turn a vague idea into a loop, or has not yet settled success criteria, boundaries, risk, or budget. First test whether a loop is justified; route one-shot or deterministic tasks to a simpler workflow. Use proportionate intake: confirm only load-bearing decisions, offer visible reversible defaults for minor details, and write durable SPEC/STANDARDS/GOALS artifacts only when the loop's scale or lifespan benefits from them. Hands settled intent to loop-engineering. Not for reviewing an existing harness (loop-review), measuring it (loop-eval), running it on a schedule (loop-ops), or post-run analysis (loop-retro).

在撰写或审阅团队周报时提供格式与内容规范。当用户说"写周报"、"检查周报格式"、 "周报该怎么写"时调用。覆盖结构、量化表达和禁忌项。不适用于日报、月度总结或 绩效自评（各有其模板）。

按《Skill 完整实操指南》执行 skill 全生命周期操作，覆盖：创建、修改加固、 测试验证、触发/执行诊断、description 与评测优化、合并拆分、删除退役。当用户要 创建/写一个 skill、改/优化/加固 skill、测一测/验证 skill 好不好用、看看 skill 为什么没触发或误触发、给 skill 写 description 或评测、合并/拆分/删除/退役 skill 时调用 —— trigger on creating, writing, editing, improving, testing, verifying, reviewing, debugging, merging, splitting, deleting, or retiring any Claude/agent skill (SKILL.md), incl. description optimization and evals. 本 skill 是 skill 工程的唯一方法论前门（writing-skills 已并入；跑基准/评测 查看器/打包时调用 skill-creator 的脚本机器）。不适用于：调用某个已有 skill 去完成它的业务任务；撰写 loop/harness 提示词本身（用 loop-engineering 组）。