debugging

When creating, modifying, reviewing, splitting, or merging a skill, maintain its entry point, structure, rule layering, and review discipline.

写 skill、改 skill、review skill、拆分或合并 skill、把稳定流程沉淀为 skill 时使用；只处理 skill 本身的入口、结构、规则分层与 review，不用于执行 skill 所描述的业务任务本身。

When there is a bug, error, failed test, regression, performance drop, unexpected result, or anomalous data, establish a tight feedback loop and find the root…

When a research design, analysis, outline, revision strategy, or other trade-off needs user alignment, converge one decision at a time with explicit user…

When assigning, supervising, reusing, restarting, or verifying a sub-agent, choose delegation by context cost, brief it precisely, watch artifacts, and verify…

Evidence before opinion: when requirements or context are unclear, or advice is needed, verify primary sources before making a judgment.

Explicitly invoked when the user asks to hand the current session to a later session; compress goals, decisions, progress, artifacts, and next steps into a…

When the user changes direction, requests a decision, restates a preference, or the agent is about to change its position, separate evidence, correction…

Explicitly invoked only when the user asks to adversarially test a named plan, claim, draft, or decision; attack its premises with evidence and report what…

委派与协调：准备把活交给 sub-agent、决定续用已有 worker 还是重新开始、已有 sub-agent 在跑需要督办、或它有返回需要核实时使用。不涉及委派、全程自己动手的执行循 executing-plans。

求真校准：当用户改变方向、要求做决策、重申其倾向，或 agent 准备随用户改变判断时自动触发；检查立场变化是否由新证据或原推理纠错驱动，并据实质疑用户。普通事实查证用 grounding，用户取舍收敛用 grilling，显式对抗审视用 red-teaming。

Before starting a new research question, analysis, chapter, or idea, expand the candidate hypotheses, research questions, and explanatory angles before…