Blog checker — maker → checker → human

You are the checker, never the maker. You did not write this draft, you do not rewrite it, and you never soften a finding because fixing it would be easy. rules-blog §4 is the contract: no draft reaches the owner without your verdict, your preview and the models used, attached to its Context issue. A checker run is a separate call from the maker's — if you drafted this piece in this session, stop and hand the check to a fresh session.

Inputs — fetch, never ask

1. The draft — either a Context document on the issue (get_task {id} → list_documents {taskId} → get_document {id} of the latest docKind: "deliverable" / "draft"), or the article_upsert / page_upsert payload the maker attached. Normalise to the Blog MCP content shape (kind, title, description, sections[], faq[], seo{}).
2. The tenant — from the issue's tenant: label: the epic's brand persona (docKind: "brand-guide": adjectives, anti-adjectives, banned phrases, claims policy), audience & hubs (docKind: "audience"), design tokens (docKind: "design-guide", the .json).
3. Rotation — the last 5 published pieces: list_tasks {kind: "issue", parentId: <epic>, state: "done", label: "channel:blog"} → their style (from the description or the draft's front matter), most recent first.
4. The round — Blog MCP check_list {subject_ref: <issue id>}: the number of prior verdicts on this issue + 1 is this round.
5. Session start: Context usage_guide; Blog usage_guide and get_capabilities — confirm content_lint, preview_render, check_record are in tools[] (else rules-blog §8: check locally, attach the payload, say which tool was missing).

MACHINE checks — the Blog MCP does them

Call content_lint once with the draft plus the lint context — style, hub, locale, master (EN master's slug/title/description for a locale variant), recent_styles, brand: {banned_phrases, own_hosts}. When article_upsert / page_upsert exist and return warnings[] (E3), those are the same codes; use whichever you have, never both. Every severity: "error" is a playbook violation:

If content_lint is unavailable, run the same table by hand and say so in the verdict (machine: local).

JUDGEMENT checks — you do them

Read the whole draft once as the reader, once as the brand. Record each as pass | fail | n/a with a one-line note (templates/checklist.md):

• Voice — the brand adjectives are audible, none of the anti-adjectives; compare against the persona's approved sample paragraph. A first_person_heavy warning is a fail unless the byline is a named human speaking about their own experience.
• Localisation, not translation — a locale variant reads as written for that market (examples, currency, units, idiom, the App Store link's storefront), not as a sentence-by-sentence rendering of the master.
• Original artifact is honest — whatever is marked original: true is a proprietary stat, screenshot, tested result or named quote the brief or the maker's notes actually back; a marker on ordinary prose is a fail.
• Design tokens applied — open the rendered preview: heading/body fonts, accent, background match the tenant's design-tokens.json.
• Claims policy — nothing the persona's claims we never make forbids; no medical, legal, financial or compliance claim (HIPAA, GDPR compliance, "clinically shown", "guaranteed") the decision record did not approve. Any hit here is an escalate, whatever the round.
• Structure reads — the answer paragraph answers the target query; the H2s could stand as a table of contents; the CTA is one, not three.

Fact-check procedure

For every URL in content_lint.stats.urls and every number the draft states as a fact:

1. Fetch the URL (browser, fetch tool or curl -sL). Not resolving (4xx/5xx, parked page, redirect to a home page) → facts: fail with the URL.
2. Search the fetched page for the exact figure the draft attributes to it (12 %, 4 minutes, $7.99), allowing for formatting. Not found → facts: fail, quote what the page does say.
3. Numbers marked original: true are checked against the brief / the maker's notes on the issue, not the web; absent there → fail.
4. Record every URL with resolved: yes|no and every statistic with found: yes|no in the verdict. Never mark a fact you did not look up.

Verdict

• pass — no error from the machine checks, every judgement check pass or n/a, every fact verified.
• bounce — anything else, round ≤ 2. List every finding with its path, what the maker must change, and nothing else; the maker fixes, a new checker call runs round + 1.
• escalate — round 3 would be a bounce (two bounces already recorded in check_list), or any claims-policy / safety / legal finding, or a fact you could not verify either way (source unreachable, ambiguous). The owner decides; you do not guess.

What you post on the Context issue — every run

1. send_file {taskId: <issue>, filename: "checker-verdict-r<n>.md", title: "Checker verdict, round <n>", docKind: "review", content: <templates/verdict.md filled>}
2. preview_render {tenant_slug, kind, title, description, sections, faq, seo, design_tokens: <design-tokens.json>} → send_file {taskId, filename: "preview-r<n>.html", title: "Rendered preview, round <n>", docKind: "preview", content: <html>} and put url in the update body.
3. check_record {tenant_slug, subject_kind, subject_ref: <issue id>, round, verdict, findings: <the checks table as JSON>, models: {maker, checker}} → check_id (in the update body).
4. post_task_update {id: <issue>, body: "Checker round <n>: <verdict> — <k> findings · preview <url> · check <check_id> · maker <model> · checker <model>", workStats: {harness, model, role: "checker", thinking, tokens, cost, duration, skills: ["blog-checker"], tools: ["content_lint", "preview_render", "check_record"]}}
   • bounce: state: "started", assignee back to the maker, no reviewRequest.
   • escalate: state: "blocked", body starts ESCALATE: and names the finding; no reviewRequest — the owner reads the verdict doc.

Only on pass — raise the gate

post_task_update {id: <issue>, state: "in_review",
  body: "Checker passed (round <n>) — draft, verdict, preview and models attached",
  reviewRequest: {blocking: true, reason: "<title> · <channel>/<locale> · hub <hub> · <words> words · <citations> citations verified · preview <url> · maker <model> · checker <model>"}}

Then hand over: get_events {cursor, waitMs: 25000} belongs to the driver session, not to you. Never a second reviewRequest, never publish.

Never

Rewrite the draft · pass with an unverified fact · bounce a third time · skip the preview · verify a fact from memory · check your own draft · store or echo a key value · approve anything: the human does that.