Infinite run ai seo

Infinite Run AI SEO

A persistent, backlog-driven loop that keeps improving a site's AI-search performance forever. Each tick does one sensible unit of work, then re-arms. It does NOT deploy every tick. It does NOT fire on a 6s cadence.

Why this exists

A fixed 6s loop floods notifications and pushes empty commits to Vercel. A real AI-SEO cycle (audit → improve → build → deploy → propagate) takes minutes. This skill makes the loop resumable (state on disk) and self-expanding (it finds new queries), so "run forever" is safe.

Files (per site)

All loop state lives in the site repo so it is inspectable, commitable, and survives stop/restart.

<site-repo>/.ai-seo/
├── backlog.jsonl      # one query per line, the loop's memory
├── runs.log           # append-only tick log (ts, action, outcome)
└── config.json        # cadence, deploy gate, discovery interval

backlog.jsonl

One JSON object per line. Status: pending → working → done | skip.

{"q":"apakah ada kelas AI untuk anak di Indonesia","page":"/","status":"done","commit":"3a7c7a9","ts":"2026-07-01"}
{"q":"kursus AI anak SD online","page":null,"status":"pending","found":"2026-07-01","source":"discovery"}
{"q":"biaya les AI untuk anak","page":null,"status":"skip","reason":"pricing claim — off-site only","ts":"2026-07-01"}

Normalize q for dedupe (lowercase, trim, collapse spaces). Never keep two lines with the same normalized query.

config.json

{
  "cadenceSeconds": 900,
  "minDeployGapSeconds": 600,
  "discoveryEveryTicks": 10,
  "maxEmptyDiscoveryStops": 3,
  "honestyGuardrails": true
}

Tick workflow

On each wake (sentinel AGENT_LOOP_TICK_aiseo or dynamic heartbeat):

1. Read .ai-seo/backlog.jsonl and .ai-seo/config.json.
2. Pick the top entry with status=pending. If none:
   • If ticksSinceDiscovery >= discoveryEveryTicks, run Discovery (below) and re-pick. If discovery returned nothing, increment consecutiveEmptyDiscovery; if it reaches maxEmptyDiscoveryStops, pause the loop and tell the user on-site surface is saturated (off-site levers remain: indexing, Google Business Profile, backlinks).
   • Else re-arm a long heartbeat and return.
3. Audit the chosen query against the site:
   • Which page should answer it? Existing page, or does it need a new section/page?
   • Is there an extractable answer block (40–60 words, self-contained, affirmative, no foil "X not Y" phrasing)?
   • Is there matching schema (FAQPage, Course/Service, speakable)?
   • Does the page have a stable id anchor for the answer?
4. Improve: make the smallest change that closes the gap. One page, one edit. Add/strengthen the answer block, schema, or anchor. Respect the site's CLAUDE.md honesty guardrails — no fabricated testimonials, no invented prices, no unbacked superlatives. Never claim a client that doesn't exist.
5. Build (pnpm build / npm run build). If red, fix or revert; do not deploy a broken build.
6. Deploy gate — push only if ALL hold:
   • build is green,
   • the change is meaningful (not whitespace/rename-only),
   • now - lastDeployTs >= minDeployGapSeconds (anti Vercel/git spam),
   • there are uncommitted changes worth shipping. Otherwise stage the edit locally and re-arm without deploying.
7. Record: update the backlog line to done (with commit hash + ts) or skip (with reason), append a line to runs.log, update ticksSinceDiscovery and lastDeployTs.
8. Re-arm the next wake using the loop skill's dynamic schedule (fallback heartbeat = cadenceSeconds). Do NOT use a fixed 6s loop.

Discovery

Find new candidate queries to append to the backlog. Sources, in priority:

1. Related searches / "People also ask" for queries already in the backlog (variant phrasings, long-tail, locale variants).
2. firecrawl_search with diverse framings of the topic — synonyms, audience segments, intent variants. Run several distinct framings, not one.
3. Refactor winners: for any status=done query that the site now answers well, derive its long-tail variants ("X online", "X di <kota>", "X untuk <usia>", "biaya X", "berapa lama X").

For each candidate: normalize, dedupe vs every existing backlog line (any status), and append as pending with source: "discovery" and found date. Skip candidates that would require off-site proof to answer truthfully (pricing, testimonials, rankings) — mark them skip with a reason instead of leaving them as bait.

Cadence

Use the loop skill in dynamic mode. The fallback heartbeat is cadenceSeconds (default 900s / 15m). That is the cadence that actually fits a build+deploy+propagate cycle. Never run this on a fixed sub-minute interval.

Starting

/loop /ai-seo infinite-run on <site-repo>

On first start:

1. Create .ai-seo/ and seed backlog.jsonl with the user's primary target query (status pending or done if already won) plus 3–5 obvious variants.
2. Write config.json with defaults (or values the user gave).
3. Run the tick workflow once immediately.
4. Arm the first heartbeat; confirm cadence and that the loop is self-pacing.

Stopping / pausing

• User says stop → kill any watcher/heartbeat PID, do not re-arm. Backlog stays on disk so the loop is fully resumable later.
• Auto-pause → consecutiveEmptyDiscovery >= maxEmptyDiscoveryStops: stop arming and tell the user on-site levers are saturated; list the remaining off-site actions they must do themselves.

Guardrails (always on)

• One meaningful change per tick. No empty commits. No deploy spam.
• Honesty first: never fabricate evidence. If a query can't be answered truthfully on-page, mark it skip with the reason.
• Run the project's anti-AI-slop pass on any user-facing copy before it ships (no em-dashes, no "X, not Y" foils, ground or remove unsourced claims).
• Keep the backlog the single source of truth — do not keep query state in your head across ticks.