Use when something is broken or slow on a machine or a service — "it does not work", "the site is down", "the disk is full", an error in a log — and the user wants it diagnosed, fixed or written up. Drives a step-by-step diagnosis and produces an incident report.
Resources
5Install
npx skillscat add allkindev/allkin-exchange/incident-report Install via the SkillsCat registry.
SKILL.md
Incident diagnosis and report
Fix the right thing, once, and leave a trace. The method is the same whether
you can only observe (read_system) or also act (run_command): observe,
form one hypothesis, test it, then act — never act on a guess.
Procedure
- Freeze the symptom. Write down what is observed, since when, what
changed last (deploy, update, config, reboot), and what still works. Ask
the user only what you cannot observe yourself. - Observe before touching. In this order, stopping as soon as the cause
is clear: service status and its last log lines; disk, memory, load;
recent errors in system logs; network (ports listening, DNS, certificates);
recent changes (package updates, config files modified lately). - One hypothesis at a time. State it ("the disk is full, so the service
cannot write its journal"), name the one check that confirms or refutes it,
run that check. A refuted hypothesis is noted, then the next one. - Act minimally. The smallest reversible action that restores service
comes first (restart, free space, roll back one change). Each command goes
through the approval card with its explanation and risk level; a
destructive action says what it destroys. Do not run a second fix before
checking the first. - Confirm. Re-observe the original symptom. "Should be fine now" is not
a confirmation. - Write the report as
incident-<date>-<short-name>.mdin your folder, in
the shape below, in the user's language, and give the summary and the
follow-ups in the conversation.
Shape
# Incident — <short name> — <date>
**Impact:** what was broken, for whom, how long.
**Cause:** one sentence.
## Timeline
- hh:mm — symptom noticed / reported
- hh:mm — check: … → result
- hh:mm — action: … → result
- hh:mm — service confirmed restored
## Root cause
…
## Follow-ups
- [ ] what would prevent it (owner, if known)Do not
- Do not restart "to see": a restart erases the evidence of the cause.
- Do not delete logs or caches to free space before knowing what fills the disk.
- Do not touch what you have no right to touch: without
run_command, hand the
user the exact commands and what to look for.