Three Kings Manager

Autonomous owner of ladder discipline for MIU content edits. On invocation it runs the full loop — fetch, gate-check, edit eligible pages, record, monitor — and only reports the compact outcome at the end. The check output is a control plane for deciding actions, not a deliverable. It never asks permission mid-loop; the safety gates (yarn db:backup before any production blog write) are the permission.

Operating loop

Every 3 days (read-only):

0. Reconcile any history row with liveVerification.status: "pending" before judging it. Read the authenticated API field and the normal public URL's rendered article, not script/RSC text or a cache-busting URL. If public HTML is still old, retain HOLD, skip indexing and do not judge the GSC window. If the recorded field/hash and public copy match, retain the original DB date as dbEditDate, set the history date and entry lastEdit to the first confirmed public-exposure date (conservatively today if the earlier date is unknown), and set liveVerification.status: "verified" with confirmation date. Then require 14 complete post-confirmation GSC days. API success alone is not evidence of search exposure; do not deploy unrelated dirty work or purge site-wide caches to force a one-page change.
1. Refresh the correlator export (final GSC data, 3-day holdback):

   node ./.claude/skills/gsc-causation-correlation/scripts/gsc-daily-correlate.cjs --output=/tmp/gsc-daily.json
   

2. Run the check:

   node ./.claude/skills/three-kings-manager/scripts/tkm.cjs check \
     --gsc=/tmp/gsc-daily.json --ledger=docs/SEO/maintenance/three-kings-ledger.json
   

3. Act on the verdicts:

Verdict	Meaning	Action
REGRESSION	latest judged edit lost clicks or degraded position	Rollback recommended — see below
EDIT_NOW	window closed, gates pass	Execute the next rung, then record-edit
VERDICT	all 3 rungs complete	Judge with gsc-causation-correlation pre/post windows; restart ladder only with new evidence
HOLD	14-day window still open	Nothing. This is the system working
GATED	demand change / phantom cluster / junk-position bloat	No edit; inspect SERP or reclassify
STOP	stop rule fired	No edits; quarantine or consolidation evaluation only

Executing a rung (only on EDIT_NOW) — do this immediately, do not stop to report:

EDIT_NOW also requires a real CTR/snippet gap inside the composed Three Kings striking-distance band (position 5–15). The checker gates pages outside that band and pages already meeting the composed serp-ctr-snippet-rewrite-technique rank-band CTR threshold; a closed window alone is not permission to rewrite a healthy snippet or a page whose real problem is ranking/intent.

1. Rung 1 = seo_title, rung 2 = seo_description, rung 3 = proof-led body pass (direct answer + evidence module above the fold).
2. Draft the replacement copy with seo-content-3-kings-technique (its Three Kings audit + competitive-gap rules define what good copy looks like).
3. Execute the edit with the blog-edit skill (.agents/skills/blog-edit/SKILL.md) — it owns the production mechanics: PATCH /api/blog/posts/[slug] with x-api-key: BLOG_API_KEY from .env.api, GET readback, and the live-HTML spot check. Blog DB writes are effective immediately; repo/backlog changes ride the next deploy.
4. Production blog writes require the yarn db:backup gate first (project rule).
5. After the edit lands live: node tkm.cjs record-edit --ledger=docs/SEO/maintenance/three-kings-ledger.json --url=<url> --rung=<n> --note="<what changed>" --previous="<prior field value>". This re-opens the 14-day window; previous captures the old value so a REGRESSION rollback is executable without archaeology. The ladder enforces one rung at a time.
6. If API readback succeeds but public HTML remains stale, still record the exact prior field value and DB write immediately, attach liveVerification: {status: "pending", field, sha256, checkedOn, apiUpdatedAt, reason} to that history row, and HOLD it. Do not label the edit live or measure an unexposed experiment. A public-cache blocker is an incomplete end-to-end action, not a reason to fabricate success or issue another blog write.
7. Append the blog changelog entry (blog-edit after-finishing rule) and record the edit in docs/SEO/maintenance/seo-changes-backlog.md; update the GSC request-indexing backlog row and request indexing manually only after normal public HTML matches (the 3-kings skill's indexing loop rule).

Monitoring (automatic in every check): each history row is judged once its window closes — WIN (clicks/d ≥ 1.2x prior), FLAT, or LOSS (clicks/d ≤ 0.8x prior, or position degraded >2 without click gain). Outcomes persist in the ledger. A measured rollback row is diagnostic only and must never raise another REGRESSION: re-applying the already-losing experiment creates an undo loop. After a rollback window closes, route the page through the normal HOLD/GATED/STOP/VERDICT gates instead.

Rollback (executed autonomously, gated): a LOSS on the most recent edit shows as REGRESSION. First rule out demand-side losses: if the outcome row shows impDelta sharply negative (≤ -40%) while posDelta stayed within ±1, the click loss tracks a demand/SERP shift, not the edit — record the outcome, skip the revert, and treat the row like a demand-change gate. Otherwise revert immediately, with a fresh yarn db:backup first: restore the prior field value from the history row's previousValue (fallback: dated backups in backups/) via the blog API PATCH, verify with GET readback, then node tkm.cjs record-revert --url=<url> --note="<restored field>". A revert opens a fresh 14-day cooling-off window. If the field lacks a previousValue and no backup covers it, stop and surface the gap instead of guessing the old copy.

Ladder discipline

• One rung per page per 14-day window. record-edit refuses same-page edits inside an open window and refuses skipping rungs.
• After two judged zero-click rungs, set a stopRule in the ledger (manual, evidence-backed) — do not keep re-rolling the same field.
• record-revert never advances the rung.

Recorded MIU evidence

The gates and thresholds encode the 2026-09-03 causation study (see gsc-causation-correlation SKILL.md): Three Kings on pages ranking pos 5-11 compounded best-free-upscaler 5→493 clicks/wk; snippet edits on phantom clusters, demand-shift rows, and junk-position rows never earned clicks. The ledger was seeded 2026-09-03 from the backlog's real edit history, so the first check judges those edits retroactively.

Coordination

• Data: gsc-causation-correlation (daily series export consumed by check), gsc-analysis (standard 28d fetch for opportunity discovery).
• Editing: seo-content-3-kings-technique (drafting), blog-edit (execution), blog-performance-monitor (single-URL incidents).
• Reclassification: cannibalization-consolidation-technique (STOP rows, cannibalization), quarantine rules in serp-ctr-snippet-rewrite-technique.