| name | content-curator |
| description | Run one agentic, on-brand content-improvement pass for a single IT-Journey collection and open one gated PR. The Lane B (substantive) executor that wires the cms-curator loop to the brand-voice system. Use when asked to "improve a collection", "run the content factory", or driven by content-factory.yml in CI. |
You are running one substantive content-improvement pass for a single collection. This skill is deliberately thin: it composes two skills you already have — cms-curator for the loop mechanics and brand-voice for how the writing must read — and adds the per-collection, open-one-PR discipline the content-curator agent and .github/workflows/content-factory.yml rely on.
It ends in exactly one pull request (or a clean no-op with a reason), never an open-ended rewrite.
0. Inputs
COLLECTION — one of quests · docs · notes · quickstart · about (the loose
pages/ root counts as root). You improve only this collection this run.- Auth + model come from the runner (
scripts/ai/run.sh + _data/ai.yml); you
don't manage them.
1. Orient (delegate to cms-curator §1)
make cms-all
make cms-status
Open .cms/worklists/<YYYY-MM-DD>.md and read Lane B (substantive) rows whose path is in your COLLECTION. That is your candidate set. Also read the policy files cms-curator §0 lists (.cms/config.yml, the per-collection .github/instructions/*.instructions.md, _data/brand/).
2. Load the voice (delegate to brand-voice)
Run the brand-voice skill for COLLECTION so the section's profile, values, and _data/brand/glossary.yml are loaded before you write. The arc is zer0 → her0; teach, don't hype.
3. Pick exactly one target
- Dedup first:
gh pr list --state open --label auto:content --json title,headRefName.
If your best candidate already has an open PR, take the next.
- Prefer the lowest health / most drift Lane B item for the collection. If Lane
B is dry, pick the weakest existing page in the collection (lowest cms health) and improve that — never silently no-op.
4. Improve it (the real work)
Apply cms-curator §"Lane B" judgment plus the brand-voice bar:
- Tighten prose; fix structure; add a concrete "you'll know it worked when…".
- Normalize frontmatter to CI constraints (title 30–60, description 120–160,
ISO-8601 dates, YAML-list tags/categories). Reuse
make content-normalize-apply for mechanical bits — don't hand-fix what a fixer handles.
- Kill drift:
python3 scripts/ci/brand_lint.py <file> must come back clean
(fix preferred spellings and replace discouraged hype with specifics).
- Honesty rule: every command/output/link is real and verified.
5. Verify (delegate to cms-curator §validate)
make prose-oneline-apply
make content-audit
make quest-data && make quest-audit
One paragraph per line. Write body prose as a single unwrapped line per paragraph — never soft-wrap at ~80 cols. The markdown-oneline CI gate fails any PR with wrapped prose, so run make prose-oneline-apply after you edit (it only joins prose; code/tables/Liquid/front-matter are left untouched) and stage the result.
Do not open a PR that fails. Fix or shrink scope until green.
6. Open exactly one PR
git switch -c content/curator-<COLLECTION>-<short-slug>
gh pr create --label auto:content --label collection/<COLLECTION> \
--title "content(<COLLECTION>): <concise what>" \
--body "<what changed + why, the brand_lint result, and any infra fix you could NOT make (out of scope)>"
echo "<pr-url>" > pr-result.txt
Then STOP. One collection, one PR, minimal diff. Never merge; the gate and the auto-merge workflow decide. Never touch .github/**, .claude/**, scripts/**, _config*.yml, _data/brand/**, .cms/**, or dependencies — flag those in the PR body instead.
