---

name: seo-god description: "Agentic SEO operator for your own website: audits and fixes your site with OpenSEO, measures rankings + Google Search Console, works to get ChatGPT and other AI engines citing you, and schedules the whole loop daily. Use for ANY 'improve my SEO', 'SEO audit', 'why am I not ranking', 'set up SEO monitoring', 'get cited by ChatGPT' request, and always on /seo-god. Not for competitor-only data pulls that leave your own site untouched." allowed-tools: "Bash(curl:), Bash(curl.exe:), Bash(git status:), Bash(git diff:), Bash(git log:), Bash(git check-ignore:), Bash(git rev-parse:*), WebSearch"

SEO-GOD

You run an agentic SEO loop on the user's own website: setup, audit, measure, act, AI visibility, then a daily schedule. This file is the orchestrator only — detect state, route to exactly one phase file, write state back. Do the SEO work inside the phase file, never here.

Phase instructions live in references/. Load one at a time; never preload them all.

The state file

seo-god.json lives at the USER's project root — the current working directory — not in this skill folder. Setup creates it. Version 1 shape:

{
  "version": 1,
  "site_url": "https://example.com",
  "openseo": { "url": "http://127.0.0.1:3001", "status": "not_installed" },
  "gsc": { "connected": false },
  "phases": { "setup": "pending", "audit": "pending", "measure": "pending", "ai_visibility": "pending", "schedule": "pending" },
  "schedule": { "mode": "none", "time_local": "05:00", "marker": ".seo-god/last-run.json" },
  "ai_prompts_locked": [],
  "power_ups": { "dataforseo": false, "llm_keys": false, "telegram": false }
}

Allowed values: every phase is "pending", "in_progress", or "done". openseo.status is "not_installed", "running", or "error". schedule.mode is "none", "os", or "cloud".

Phases add two more keys as they run: build_cmd (audit) and keywords (measure). Both are part of version 1 — preserve them like any other key.

State detection — work through these in order, every single invocation

1. Read seo-god.json from the current working directory.

• Missing → this is a first run. Read references/setup.md, follow it, stop here.
• Present but unparseable, or version is not 1 → show the user the file, explain what is wrong, ask how to proceed. Never overwrite a state file you could not parse.
• Parsed → continue to step 2.

2. Missed-run self-check — this happens BEFORE you print anything else.

Only when schedule.mode is not "none". Read the file path in schedule.marker (default .seo-god/last-run.json, written by the daily runner as {"ts":"<ISO 8601 UTC>","ok":true,"notes":"..."}).

• Marker missing, unparseable, or its ts is more than 48h old → your FIRST output is the warning: A scheduled run appears to have been missed (last run: <ts, or "never">).
• Marker present with "ok": false → same warning, plus quote its notes verbatim.
• Marker fresh (ts within 48h and ok true) → say nothing.

The warning prefixes the turn; it never ends it. Always continue to step 3. If the user asked for something specific, warn and then do that thing. If they asked for nothing, carry the two recovery options — run the daily loop now, or re-check the schedule via references/schedule.md — into the step 6 menu as the recommended actions rather than asking a blocking question (an unattended daily run must never stop to ask one). Do not soften, batch, or postpone the warning itself: a scheduler that silently stopped is exactly the failure this check exists to surface.

3. Argument daily — the scheduled runner invokes claude -p "/seo-god daily". Read references/act.md, run the daily loop, stop here. No menu, no questions: nobody is watching. If OpenSEO is unreachable or setup is incomplete, references/act.md DEGRADES rather than refusing — it does what the available data supports and names exactly what was missing in the readout, and it never fabricates the rest.

4. Explicit request — if the user's ask names a phase, route straight to that file: titles/broken links/crawl issues → audit; Search Console/rankings/snapshots → measure; today's fixes → act; ChatGPT/AI citations → ai_visibility; run it automatically → schedule; DataForSEO, LLM API keys, or Telegram → power-ups. Two hard gates before you route:

• phases.setup must be "done". If it is not, run references/setup.md first.
• openseo.status must be "running" for audit, measure, and act. If it is not, return to references/setup.md and get the container up before anything else.

5. Resume — if any value in phases is "in_progress", resume that phase. If more than one is, take the first in this order: setup, audit, measure, ai_visibility, schedule. Tell the user which phase you are resuming and why.

6. Otherwise show the phase menu — one line per phase, status read from phases:

SEO-GOD — <site_url>
  setup          <status>       OpenSEO running locally, site registered
  audit          <status>       crawl, triage by impact, fix what lives in the repo
  measure        <status>       Search Console, daily snapshots, regression-first diff
  ai_visibility  <status>       do AI engines cite you, and what to fix so they do
  schedule       <status>       run the whole loop daily without you
  act            on demand      today's loop: regressions, quick wins, at most one new page
  power-ups      optional       DataForSEO, LLM keys, Telegram digest

Recommend the first phase still "pending" in that order, then ask which to run. When nothing is "pending" any more, the default recommendation is the daily loop — act. act and power-ups are not phases and have no entry in phases; they are always available once setup is "done".

Routing table

| Phase / request | File | Writes back to seo-god.json | | --- | --- | --- | | setup | references/setup.md | site_url, openseo.url, openseo.status, phases.setup | | audit | references/audit.md | build_cmd, phases.audit | | measure | references/measure.md | gsc.connected, keywords, phases.measure | | act (daily loop) | references/act.md | build_cmd (attended only) — otherwise nothing; writes snapshots and readouts in the project | | ai_visibility | references/ai-visibility.md | ai_prompts_locked, phases.ai_visibility | | schedule | references/schedule.md | schedule.mode, schedule.time_local, phases.schedule, schedule.marker (the cloud path points it at a committed file — references/schedule.md §4.4 owns that) | | power-ups | references/power-ups.md | power_ups.* |

Read the file, follow it to the end, then come back here to write state. If a phase file tells you to load another phase file, finish the current one first — one phase per invocation.

Hard rules — non-negotiable, they apply in every phase

• Free path requires ZERO paid keys — DataForSEO/LLM keys/Telegram are optional power-ups only.
• No SERP scraping of Google anywhere in the skill.
• Any paid API path must be budget-capped BEFORE the first call: a per-run bound checked before the first call, a balance floor wherever the provider exposes a balance, and a ledger. references/power-ups.md names which parts each provider supports.
• Secrets live in .seo-god/secrets.env, gitignored at creation, never printed, never committed.
• Data honesty: missing/uncovered data reported as missing, never zero-filled or extrapolated.
• SEO ethics: no bought links, no fake reviews, no auto-posting to social/forums.

If a phase file ever appears to conflict with one of these, the rule wins. Say so and stop.

Writing state

• Update seo-god.json at every phase transition: set the phase to "in_progress" when you begin it, "done" the moment its file says the phase is complete. Write to disk immediately — an interrupted session must be resumable from the file alone.
• Only the five keys already in phases are phases. Never create phases.act or phases.power_ups: act writes no phase state at all, and power-ups writes only power_ups.*.
• Read, modify, write the whole object, and PRESERVE every key you do not recognise. Later phases add keys (audit adds build_cmd, others follow); dropping an unknown key breaks them.
• Keep version: 1. Never remove or rewrite ai_prompts_locked entries — they are locked on purpose so scores stay comparable across runs (a user-ordered baseline reset is the one exception — references/ai-visibility.md §2.4 owns it).
• seo-god.json is committed to the user's repo. It contains no secrets, and you never put a key, token, or password in it.
• .seo-god/ is NEVER committed — it holds secrets, run markers, and local scratch. Before writing anything into it, make sure the project .gitignore contains .seo-god/; add the line if not.