agentleFS
fenjo26fenjo26/OpenGSC/.agents/skills/seo-production/SKILL.md

seo-production

Skill57 starsChanged 41 days agoMITView on GitHub

Produce a new article end to end through the OpenGSC MCP: demand evidence, an approved outline, a claim ledger, the draft, deterministic verification, and a package ready for Content Operations. For refreshing an existing page use page-optimization instead.

The file

We split this skill into the rules it gives an agent

6 rules. About 0 are common, 5 are distinctive, and 1 is very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈5

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 1

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Use page-optimization instead when the page already exists and underperforms.

    line 14
  2. Distinctive:

    Task card: Before any writing, state in five lines: target query, secondary queries, search intent, the reader, and what the page must let …

    line 40
  3. Distinctive:

    Claim ledger: Every number, date, price, statistic or named study in the outline gets a row: the claim, its source URL, and the date it was …

    line 52
  4. Distinctive:

    Read it as a gate, not a score:

    line 58
  5. Distinctive:

    Call it before any other tool.

    line 108
  6. Distinctive:

    Before any PAID tool, check the research log: if the same research ran within the last 30 days, reuse that result and say so instead of …

    line 113

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

OpenGSC SEO Production

Goal

Turn a topic into a publishable article that is grounded in this instance's own data, where every factual claim can be traced to a source the user can open. The deliverable is a content package, not a wall of prose: brief, outline, draft, claim ledger and the verification output.

Use page-optimization instead when the page already exists and underperforms. This skill is for something new.

Required inputs

  • The site. Call list_sites if the user has not named one.
  • A topic, a seed keyword, or a target page. Anything vaguer than that, ask.

OpenGSC MCP tools

  • get_keyword_demand: start here. Research already stored, joined against the site's own GSC positions, each row verdicted reach / wrong_page / none. Free.
  • get_competitor_gap: competitors' keywords bucketed close / weak / missing — what the market covers and this site does not.
  • get_search_performance and get_striking_distance: what the site already ranks for. A new article that targets an existing strength creates a cannibalization problem instead of traffic.
  • get_cannibalization: run it before writing, not after. If an existing URL already owns the intent, the honest recommendation is to extend that page.
  • get_generations: what has already been written here, so a "new" article is not a near-duplicate of one from three months ago.
  • fetch_page_content: any competing URL as clean Markdown. This is the sourcing tool.
  • analyze_text: deterministic check of the finished draft. No model call, same answer every time.
  • research_keywords: paid. Only when demand data is genuinely missing and the user asks.

Workflow

  1. Task card. Before any writing, state in five lines: target query, secondary queries, search intent, the reader, and what the page must let them do. Get it confirmed. Everything downstream is judged against this card.
  2. Demand evidence. get_keyword_demand, then get_competitor_gap. If neither shows real volume, say so plainly — "no demand data" is a finding, not a reason to invent an audience. Missing metrics mean not measured, never zero.
  3. Conflict check. get_cannibalization and get_search_performance for the target query. An existing ranking URL means: recommend extending it and stop, unless the user overrides.
  4. Source pass. fetch_page_content on the two or three URLs currently ranking. Read them for what the intent actually demands — sections, depth, format — not for sentences to reuse.
  5. Outline first. H2/H3 structure, one line per section on what it answers and which source backs it. Show the outline and wait. Writing before the outline is agreed wastes the draft.
  6. Claim ledger. Every number, date, price, statistic or named study in the outline gets a row: the claim, its source URL, and the date it was retrieved. A claim with no row does not enter the draft. This is what keeps the article defensible later.
  7. Write. Fill the approved outline. Concrete over decorative; the reader's task over the word count. Every ledger claim keeps its number exactly as sourced.
  8. Verify. analyze_text with the draft as text and the concatenated sources as source. Read it as a gate, not a score:
    • factDrift.severity: danger — numbers appear in the draft that no source contains. Fix before showing anyone. This is the check that stops an invented price from shipping.
    • structure.ok: false — the draft drifted from the approved outline.
    • machine tells — fix the specific ones reported, then re-run analyze_text. Editing for naturalness can silently reintroduce drift, so the last run must be the one you report.
  9. Package. Hand over title, meta description, slug, the article, the claim ledger and the verification output. Content Operations is where the human moves it: queue → approval → review → diff → pull request. This skill never opens a PR and never publishes.

Output format

  • The task card, one line per field.
  • Demand evidence: query, volume, current position, verdict — as a table.
  • The approved outline.
  • The article.
  • The claim ledger:
ClaimSourceRetrieved
  • The analyze_text verdict verbatim: uniqueness, fact drift, structure, tells.
  • Open questions: anything you could not source, stated as a question rather than smoothed over.

Guardrails

  • Never invent experience. No "when I tested this", no fabricated case studies, no invented customer quotes. Fake first-hand experience is the one E-E-A-T signal that turns into a liability the moment a reader checks. If the article needs experience it does not have, ask the user for it.
  • Never write around a missing source. Drop the claim or ask. A sentence engineered to sound sourced without being sourced is worse than an obvious gap.
  • Do not optimize for AI detectors. Detector-dodging rewrites damage the text and prove nothing. analyze_text reports machine tells so you can fix real habits — em-dash pile-ups, "furthermore", "it is important to note" — not so you can game a classifier.
  • Do not call the paid tools without being asked. research_keywords, start_generation_job and start_rewrite_job spend the instance owner's own credits and refuse to run without confirm: true. Never set that flag to work around the refusal. You have the brief in context — writing the draft yourself costs nothing extra.
  • Respect the editorial policy. The instance owner's style guide and banned-word list live in SEO Tools → Editorial Policy. If the user has one, ask for it and apply it; do not invent a house style.
  • One article per run. Bulk production is what the app's own paid pipeline is for, and that is the user's decision to make, not yours.

Site context — read first, free

get_site_context is this skill's step zero. It returns the site's shared memory — business overview, current goal, positioning, writing preferences, the competitor list, key pages and the research log — and grounds every decision below in what the operator already told us.

  1. Call it before any other tool. If it errors or comes back empty, continue without it; never block the workflow on missing context.
  2. This skill especially needs: writing_preferences, positioning. If a needed section is empty, do a minimal inline fill — ask the user (or infer from the site and confirm), write it back with update_site_context, then continue. Never front-load a full interview.
  3. Before any PAID tool, check the research log: if the same research ran within the last 30 days, reuse that result and say so instead of re-buying it.
  4. On finish, write back what is durable — new or corrected key pages, competitors the work revealed, positioning facts the user stated — and append a research-log entry if the session spent paid units (appendResearchLog).

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.