| name | create-runbook |
| description | Author a reusable runbook, or extract one from a completed investigation, following this repo's runbook standard. Use whenever a multi-step procedure, diagnosis, or recipe should be captured so future agents can reuse it. |
| disable-model-invocation | true |
| metadata | {"internal":true} |
Create Runbook
Capture procedural knowledge as a runbook in runbooks/, following the authoring
standard in runbooks/AGENTS.md. Read that file first — it defines the retrieval
mechanics, frontmatter contract, split/merge rules, fact-ownership model, contract
style, and validation. This skill is the procedure for producing one runbook.
When to use
- After a multi-step investigation whose pattern others could reuse.
- To add a new stage to the investigation pipeline.
- To turn an ad-hoc recipe (a query pattern, an optimization) into a durable how-to.
Steps
-
Check it should be a new file. Search the existing registry
(panda search runbooks "…" with a few phrasings). Content that would always be
retrieved together with an existing runbook belongs IN that runbook — the split
axis is co-retrieval, not topic count. Only mutually-exclusive-in-use content gets
its own file.
-
Extract the pattern, not the incident. Generalize: what is the reusable
diagnostic shape? Drop run-specific ids, timestamps, and host names.
-
Engineer the retrieval surface first. Write the description (what + when,
third person, ≤1024 chars) and 3-6 triggers as the literal queries a caller would
type, packed with exact identifiers (table names, error strings, API names). Body
sections are embedded too (per-section child vectors), so mid-document facts are
findable — but the metadata surface dominates ranking, so it decides whether the
runbook wins the queries it should.
-
Make it self-contained for its task. Restate load-bearing facts briefly with a
pointer to their canonical owner (runbooks://<owner>); reserve links for adjacent
tasks. A required fact that lives only behind a link is a bug.
-
Write to the standard's structure: owns-line, Inputs (liberal), Output,
Procedure; tables for decision matrices, numbered lists for procedures, positive
directives, decisive rule up front. If the output feeds another stage: one filled
example, ≤10 top-level fields, summary/reasoning before verdicts, confidence as
low|medium|high per runbooks://evidence_discipline.
Verify
go test ./runbooks/... — frontmatter, refs, uniqueness.panda search runbooks "<one of your triggers>" returns the new runbook.scripts/runbook-retrieval-check.sh still passes (add a golden query for the
new runbook while you're there).- The server logs
Runbook registry loaded with the new count.
$ARGUMENTS
