AgentSkillsCN

agent-observability

通过记录用户修正操作并进行后续审核，让代理自行上报问题，随后再以正确技能继续处理。当用户说“别那样做”、“停止做 X”、“以后一定要做 Y”，或主动请求自我修正时，可使用此方法。

SKILL.md
--- frontmatter
name: agent-observability
description: Self-report agent issues by logging user corrections for later review, then resume with the correct skill. Use when a user says "don’t do that", "stop doing X", "always do Y", or requests self-correction.

Agent Observability (Self-Reporting Log)

Operating rules

  • •Treat user corrections as issues to log.
  • •Log locally in the current repo at docs/observed-coding-agent-issues.md.
  • •Create the report file if it doesn’t exist.
  • •Mandatory routing checklist (do not skip): (1) log the correction, (2) read the relevant skill file, (3) respond to the main task. This skill never handles the main task directly.
  • •After logging, always open the relevant skill file before continuing, even for small/obvious tasks.
  • •If the user request includes a concrete edit/implementation (README/docs/config/code), open skills/coding/SKILL.md before any questions, reads, or edits.
  • •If the user asks for a plan/spec, open skills/planning/SKILL.md before drafting the plan.
  • •If the user asks to create/update a skill or references skills//SKILL.md, open skills/skill-creator/SKILL.md before continuing.
  • •On multi-turn requests, re-evaluate routing at the start of each turn; logging does not replace skill invocation.
  • •When opening references, use full repo paths like skills/agent-observability/references/... (not references/...).
  • •When a reference trigger clearly matches, open the referenced file before drafting output.
  • •If the user provides a word/length limit, keep the log entry short and minimize extra reads.
  • •Rate limit: at most one log entry per session unless the user explicitly asks to log more.
  • •Batch multiple corrections into a single entry by updating the existing root cause and guardrail notes; do not add verbatim quotes.

Detect → Record → Resume

1) Detect correction

Trigger on phrases like or similar to:

  • •“don’t do that”, “stop doing X”, “always do Y”, “never do Z” Friction alone is not sufficient; require a correction or explicit request.

2) Record issue (append)

Append a short entry to docs/observed-coding-agent-issues.md. If a log has already been written this session, batch new corrections under the same entry unless the user explicitly asks for another log.

Entry format:

  • •Date/Time:
  • •Task/Context:
  • •What went wrong (root cause):
  • •Proposed guardrail:

3) Resume main task

Re-evaluate which skill applies next (coding/skill-creator/planning) and open that SKILL.md before responding. Do not continue the task until the relevant skill is loaded.

Routing shortcuts (read immediately)

  • •README/docs/config/code edits → read skills/coding/SKILL.md
  • •Plan/spec request → read skills/planning/SKILL.md
  • •Skill creation/update or skills//SKILL.md mention → read skills/skill-creator/SKILL.md
  • •If correction + task appear in the same prompt, still read the task skill before any tool calls.

If unable to write

  • •Output a “Self-report draft” block in the response with the entry content.
  • •State that it was not written and needs manual application.

Reference triggers (open when clearly relevant)

  • •Policy change / PR summary requested -> skills/agent-observability/references/pr-template.md

References

  • •skills/agent-observability/references/pr-template.md - PR summary template for policy changes