| name | docs-sync |
| description | Keep the docs current after a code change — map changed code areas to the docs that describe them, update those docs in place, and prune known-issues rows the code has verifiably fixed. Invoke after editing the pipeline, adding a model/agent/task/metric, moving a directory, or whenever someone asks to "sync the docs", "update the docs for this change", or "do the docs still match the code?". |
Sync docs to a code change
Code moved; the docs should move with it. This skill takes a change set and walks
it back to the docs that describe the touched areas, updates those docs, and
removes stale recovery rows the code now fixes. It edits prose only — it does not
change behavior.
Follow the repo's doc conventions while editing: GFM, humanized (not journal-like
or changelog-y), minimal, no contradictions or duplication, and no model
scores or result tallies in any doc.
Flow
1. Get the change set
Take the diff or the list of changed files (e.g. git diff --name-only main...,
or the working tree). Group the changes by area — a model provider, a deployer, an
agent harness, the CLI, metrics, a task, the leaderboard, a directory move.
2. Learn where things are documented
Tree-walk docs/ and read the doc map in
../../../docs/README.md. Also read the nearest
AGENTS.md to the changed code — the root AGENTS.md, plus
devops_bench/AGENTS.md, site/AGENTS.md, or tasks/AGENTS.md — since those
route contributors and often need the same edit as the docs.
3. Map changed area → affected docs
A change can touch more than one row (e.g. a new CLI flag that exposes a new
metric). Update every doc the change actually affects, and only those.
4. Update the affected docs
Edit in place. Keep edits surgical — change the lines the code change made wrong,
don't rewrite the page. Match the page's existing voice and structure. If a code
change makes a documented example wrong, fix the example. Never paste a score
table or run results into a doc.
5. Prune the known-issues router
This is the inverse of what the run skills do: they append freshly observed
failures to known_issues.md; this
skill removes rows the code now fixes.
For every Section 1 (router) or Section 2 (known hacks) row whose root cause your
change set addresses:
- Confirm the fix exists in the code before removing the row. Open the file
the row points at and verify the fix is actually there — cite the file and line
in your report. A row describes a current failure mode; remove it only when
that failure mode can no longer happen.
- If a hack's stated removal condition is now met, remove its Section 2 row.
- If you're not certain the fix is complete, leave the row and say so — a stale
"still broken" row is safer than a wrong "all clear".
6. Verify links + report
Confirm every relative link you touched still resolves (open the target, or check
the path exists). Then report: the change-set → doc map you applied, each file
edited, and each known-issues row removed with the file/line that proves the
fix.
Wrong tool?
- A new failure to record (not a fix) → that belongs in the run skills, which
append to the known-issues router; this skill only prunes verified fixes.
- Authoring brand-new docs for a feature with no home yet → write the page,
then add it to the doc map in
../../../docs/README.md
and the nearest AGENTS.md.
Guardrails
- No model scores, leaderboard standings, or result tallies in any doc — ever.
- Never remove a known-issues row without citing the code that fixes it.
- Edit prose only; this skill does not change code behavior.
- Keep edits minimal and humanized; don't turn a doc into a changelog.