weekly-audit

Companion skill for @automatelab/seo-performance-mcp. Where the base seo-performance skill routes a single question to the right tool, this skill runs the full weekly playbook end-to-end and returns one ranked action list.

The MCP server already exposes an audit_cohort prompt that runs a cohort report and emits briefs. This skill is one layer higher: it cross-references three independent signals (quick wins, cohort verdicts, citation losses), dedupes by URL, and ranks by how many signals agree.

When to use

The user wants the answer to "what should I edit this week" in one go - not a chat-driven exploration. For single-URL questions, defer to the base seo-performance skill.

What it does

Read-only. Proposes edits. Never applies them. Wiring the apply path is up to your CMS - keep audit and apply as separate skills.

Steps

1. Pull three lists in parallel

Issue these in a single message:

• gsc_quick_wins with window=90, min_position=5, max_position=15, min_impressions=50, limit=20
• cohort_report with min_age_days=90, window=30, limit=20
• posts_list with limit=50, min_age_days=30 (input for the citation sweep)

2. Citation-loss sweep

For the top 15 URLs from posts_list (oldest first), call posts_cite_loss per URL. Keep only URLs with losses[].length > 0. If CITATION_INTELLIGENCE_URL is unset, this step returns empty - skip it.

3. Dedupe and rank

Merge by URL. A URL appearing in multiple lists gets a priority boost:

Sort by total weight desc, then by cohort_report confidence desc. Take top 5.

4. Per top-5 URL, propose one concrete edit

Pick the single highest-lift edit for the dominant signal:

• Quick-win URL with 0% CTR query at position 5-15 -> rewrite meta_title using the query verbatim, under 60 chars.
• `refresh` verdict with `decay_30d_over_30pct` -> propose a new H2 + intro paragraph targeting the top GSC query.
• `merge` verdict -> name the sibling URL and recommend a 301 target.
• Citation loss -> propose H1 + lead-paragraph phrasing that mirrors the lost query verbatim (LLMs cite phrases, not paraphrases).
• `expand` verdict -> name 3 FAQ questions to add (pull from posts_snapshot top queries).

5. Output format

Markdown digest. For each of the 5:

## <n>. <URL> -> <verdict label or "quick-win">

**Why:** <one-sentence plain-English reason>

**Numbers:** clicks <X>, impressions <Y>, avg pos <Z>, top query "<q>" (<CTR>% at pos <P>)

**Proposed edit:** <verbatim copy of the new meta_title / H1 / FAQ block>

End with a one-line summary: "Top 3 to ship this week: <url1>, <url2>, <url3>."

House-style rules - fill in for your site

The MCP doesn't know your brand voice. Before proposing edits, the skill should enforce your rules. Drop a list here when adopting the skill, e.g.:

• Title case vs sentence case
• Em-dash policy (allowed / replace with hyphen / replace with comma)
• Emoji policy (allowed in titles / body only / banned)
• Canonical URL shape (e.g. /blog/<slug>/ vs /<slug>/)
• Forbidden words or phrases
• Persona / voice notes (plain words, no jargon, etc.)

If a proposed edit violates a rule, rewrite it before showing.

What this skill does NOT do

• It does not apply edits. No CMS writes. The output is a digest you paste into your CMS or hand to a future apply-skill.
• It does not duplicate the MCP. Tool contracts live in the MCP server.
• It does not auto-publish, auto-tag, or auto-redirect.

Known limitations

• Sitemap `<lastmod>` is unreliable if a recent deploy reset all values. For accurate age filtering, configure Ghost (or POSTS_LIST) so per-URL getPostMeta has a real source.
• Verdict engine needs >=90-day-old posts to fire useful verdicts. New sites get mostly hold/fresh_post_too_young. Lean on gsc_quick_wins (no age filter) until the cohort matures.
• Citation-loss step is a no-op without `CITATION_INTELLIGENCE_URL`. Set it when ready; until then the sweep contributes nothing to ranking.