curl -fsSL https://agents.onecontext.me/skills/blog-assessment.md -o .claude/skills/blog-assessment/SKILL.md

Blog assessment — weekly, agent-side

You are the assessor. Once a week you say how the blog is doing with real numbers, which posts AI engines cite and which they do not, and what to write or refresh next. Context stays thin: the Blog MCP only stores what the nightly cron ingested and what you ingest; the probe runs here, on this host, with the owner's keys by name. You never invent a number — a source that is not wired is reported as not configured: <NAMES>.

Inputs — fetch, never ask

1. Session start — Context start_context; Blog MCP usage_guide and get_capabilities. Confirm stats_summary, visibility_report, visibility_results_ingest, tracking_setup_status are in tools[] (contract ≥ blog-1.3.0). Missing → rules-blog §8: say "not available yet on this server", report from Context alone, stop after the report.
2. The tenant — tenant_slug from the routine prompt / the epic's tenant: label; tenant_get for hostname, locales, hubs.
3. Setup state — tracking_setup_status { tenant_slug }. Its missing[] are NAMES; copy them verbatim into the report's "Not configured" line. Do not try to fix them.
4. The lanes — from the tenant epic (get_epic {id: <epic>} → children): the Performance Report issue (lane:performance), the Topic Lane (lane:topic), the AEO/SEO Health issue (lane:aeo-seo).
5. The query list — visibility_report { tenant_slug, period: "7d" } → active_queries. If 0: build ~30 from the Audience & hubs document (docKind: "audience": target queries per hub, locale variants for the tenant's full locales) and visibility_queries_set them once, with topic = hub slug and article_id where a post targets the query.
6. Host capabilities — which key NAMES exist here (vault get OPENROUTER_API_KEY succeeds / fails; same for PERPLEXITY_API_KEY, SERPAPI_KEY). Never print a value. No OPENROUTER_API_KEY → skip the probe entirely (§ Probe), report says so.

Step 1 — stats (read-only, Blog MCP)

Call, all with period: "7d" and once with "28d" for context:

Call	Use in the report
stats_summary	totals + delta vs previous week, channels, ai_share, ai_engines, search_console (or "not configured"), top_posts, data_sources (the last ingested day — if older than 2 days, flag "ingestion stale since ")
top_posts { limit: 10 }	the table of posts with pageviews, AI pageviews, search clicks
ai_referrals	per-engine visitors and landing pages; keep the under-count caveat
search_queries { striking_distance_only: true, limit: 20 }	refresh candidates (position 5–20); empty + reason when GSC is not wired
indexing_status { limit: 10 }	last IndexNow pings vs last publishes — every publish should have a matching indexnow with ok: true; a failed or missing ping is a Health finding
Context list_issues {parent_id: <epic>, includeClosed: true}	what published / bounced / waited this week (from the issues, channel: labels, "Published:" comments)

Step 2 — probe (agent-side, cost-capped)

Only on a host that holds OPENROUTER_API_KEY. Engines and their key names:

Engine	Needs	Model (override with --models)
chatgpt	OPENROUTER_API_KEY	an OpenAI model on OpenRouter with web search
claude	OPENROUTER_API_KEY	an Anthropic model on OpenRouter with web search
gemini	OPENROUTER_API_KEY	a Google model on OpenRouter with web search
perplexity	PERPLEXITY_API_KEY (else via OpenRouter perplexity/sonar)	sonar
google_aio	SERPAPI_KEY	Google AI Overview via SerpApi, gl/hl per market

Run the shipped script — it does the calls, the cost cap and the payload:

export OPENROUTER_API_KEY="$(vault get OPENROUTER_API_KEY)"          # never echo it
export PERPLEXITY_API_KEY="$(vault get PERPLEXITY_API_KEY 2>/dev/null)"   # optional
export SERPAPI_KEY="$(vault get SERPAPI_KEY 2>/dev/null)"                 # optional
node scripts/probe.mjs --queries queries.json --out results.json \
  --engines chatgpt,claude,gemini,perplexity --markets us --max-usd 3 --samples 1

queries.json is the visibility_report.queries[] / visibility_queries_set output (only active rows are probed; priority 1 first). The script writes results.json = the exact visibility_results_ingest body minus tenant_slug: run { trigger: "weekly", engines, cost_usd, started_at, finished_at, status, notes } and results[] with query, engine, model_slug, market, answer_text, citations[{url,title}], cost_usd, probed_at, plus status: "error" / "no_ai_answer" rows where a call failed or Google showed no AI Overview. It stops at --max-usd (default 3) and marks the run partial.

Rules: identical prompt across engines (no brand in the prompt), temperature 0, one query per call, weekly is enough (daily only for priority: 1 queries when the owner asked). Never probe more than once per run per query × engine unless --samples 3 was requested for a "reliable read". Cost cap per tenant per run: 3 USD unless the decision record says otherwise.

Then ingest:

visibility_results_ingest { tenant_slug, run: <results.json.run>, results: <results.json.results> }

The server decides cited / mentioned / absent from the tenant's owned domains and aliases and returns cited_rate; read it back with visibility_report { period: "7d" } (this week) — it also carries previous_status per query × engine (last week) and share_of_voice.

If the probe cannot run (no key, host without shell), the report's visibility section is the line AI visibility: not configured: OPENROUTER_API_KEY (plus the other missing names) and last week's visibility_report numbers, labelled as such.

Step 3 — the report (Context)

1. Fill templates/weekly.md — every number from Step 1–2, every missing source by name, no estimates. Sections: Headline · Traffic · AI referrals · Search · AI visibility (cited / mentioned / absent per engine; the posts cited; the queries where competitors are cited and we are not → share of voice) · Indexing · Health · Actions.
2. Create one sub-issue under the Performance Report lane: create_issues {parent_id: <Performance Report issue id>, issues: [{title: "Weekly report — <tenant> — <ISO week>", labels: ["tenant:<slug>", "lane:performance", "report:weekly", "gate:none"], state: "done"}]} — it is a record, not work for the owner. Attach the report: attach_artifact {parent_id: <sub-issue>, filename: "weekly-report-<yyyy-Www>.md", title: "Weekly report <week>", docKind: "report", content}. Put the headline (3 lines) in the sub-issue description and in one post_comment on the Performance Report lane.
3. AEO/SEO Health — one comment on the Health issue only when something needs the owner: ingestion stale, IndexNow failing, a site-wide impressions cliff (kill-switch: pause net-new, say so), a claims-policy risk you noticed in a cited answer.
4. Topic actions → Topic Lane — one comment listing, from the evidence:
   • refresh: posts cited last week and absent this week, striking-distance queries whose post is ≥ 13 weeks old → propose kind:refresh issues by title (do not create them unless the decision record delegates it);
   • new: active queries where no owned URL is cited on any engine and no post targets them → propose a topic per hub with the query as target_query, and topics_upsert them as status: "researched" on the Blog MCP (that is the research backlog, not a publish decision);
   • stop: queries absent everywhere for 4 weeks with no post → propose deactivating.
5. workStats on the sub-issue; models used in the body; the run's cost_usd in the report.

Degrade gracefully

Missing	Do
Blog MCP analytics tools absent	report from Context only; title suffix "(no platform stats)"; say which tools were missing
not_configured names from tracking_setup_status	copy them into the report; do not guess numbers; do not create issues to fix env
OPENROUTER_API_KEY	skip the probe; visibility = last ingested run, labelled "last run "
PERPLEXITY_API_KEY / SERPAPI_KEY	probe without those engines; list them under "not configured"
ingestion stale (> 2 days)	Health comment; numbers still reported with the last day named
no active queries and no Audience doc	Health comment asking the owner for target queries; skip the probe

Never

Invent or extrapolate a number · print a key value · probe with a brand-biasing prompt · exceed the cost cap · create publish issues or change the schedule · approve anything · run the probe on a host without keys and call it "no citations" — absence of a probe is not absence of citations.