AGENT SKILL

postmortem

Writing up an incident after it is resolved — assembling the timeline from logs and messages, separating what happened from why, and producing follow-up actions that name an owner and a trigger rather than an intention.

License 许可
MIT
最近更新
2026-11-15
Stars
★ 0
Forks
1

01 前言 Front Matter

name
postmortem
description
Writing up an incident after it is resolved — assembling the timeline from logs and messages, separating what happened from why, and producing follow-up actions that name an owner and a trigger rather than an intention.

03 SKILL.md 正文

Postmortem

When to use

A task asks you to write up an incident, outage or regression that has already been resolved, or to turn a pile of logs and chat messages into an account somebody can act on.

Not for triaging something still happening. That wants the shortest path to a mitigation, and this asks you to slow down and reconstruct.

Where this one lives

This skill sits in skills/incident/ rather than directly under skills/, which is the only structural thing it is here to show. A folder under the skills root is a source, and incident is one — so this skill can be written incident::postmortem as well as postmortem.

Both spellings resolve, and the short one is what the agent is offered while the name is unique. The qualifier exists for the day it is not: two parties may each ship a postmortem, and then naming it bare is refused rather than guessed — which turns a working grant into a loud error instead of quietly changing which procedure the caller gets.

One folder of grouping and no more. Tools and subagents nest as deep as you like; skills do not, because the agent reads them off the filesystem itself and looks exactly one level down. skills/incident/deep/postmortem/ would be invisible, and kingfisher list reports anything hiding below that rather than leaving the catalogue looking empty.

The procedure

Build the timeline first, and only then explain it. Work from timestamps you can point at — deploy records, log lines, message times. Write each entry as time, what was observed, who observed it. Resist writing causes into this section; an incident where the first theory was wrong is the ordinary case, and a timeline that already assumes the answer hides the minutes where nobody knew it.

Separate the trigger from the cause. The trigger is what happened on the day; the cause is the condition that made the trigger sufficient. A deploy that exposed a latent limit is one trigger and one cause, and a writeup that names only the deploy leaves the limit in place for the next one.

Say what detection actually was. Whether the alert fired, or whether a person noticed. Time-to-detect is the number most writeups round in their own favour, and it is the one that predicts the next incident's length.

Write actions with an owner and a trigger, not an intention. "Add a limit alert, owned by the platform team, before the next capacity change" is an action. "Improve monitoring" is a sentiment. An action nobody is named on is one nobody does.

Name what went right. A rollback that worked, a runbook that was accurate, a dashboard that showed the problem. These are the parts worth protecting when somebody later proposes removing them.

What to hand back

A timeline, a trigger, a cause, detection and recovery times, and the actions. Say plainly where the record is thin — an unlogged window is a finding about the logging, and an account that quietly smooths over it teaches nobody anything.

04 同仓库更多 Skills