google-ads-routine

Google Ads — Routine (cadence + decision gates)

This is the operating rhythm, not a new analysis engine: it calls tracker and optimizer at the right cadence, remembers history, and gates actions so nothing changes unless its preconditions are met. It answers: "what do I check today, what am I allowed to change, and under what conditions."

STEP 0 — Read state + build today's agenda

1. Load account-context.yaml (esp. guardrails, margin_tiers, campaign_defaults).
2. Run the history engine: python "${CLAUDE_PLUGIN_ROOT}/skills/routine/scripts/routine_state.py" [--today YYYY-MM-DD] It reads routine-state.json in the working dir and prints what's DUE / OVERDUE per cadence.
3. Run everything that's due or overdue — not just what the user asked. If the user says "daily" but the weekly/monthly is overdue (they haven't checked in a while), run those too and say so. This is the "it remembers and nags" behavior: long gap → the overdue cadence surfaces automatically.

Model dispatch (run cheap, decide expensive) — see ${CLAUDE_PLUGIN_ROOT}/references/model-tier-dispatch.md

• Scout (haiku) — STEP 0 routine_state.py run (agenda) and the --mark stamp after acting.
• Routine / Judge — inherited from the skills this orchestrates: tracker and optimizer carry their own per-STEP tiers (collection is Routine, verdicts are Judge).
• Judge (main session) — the decision gates (every precondition check, +10% vs WAIT), reading the cooldown from guardrails, and deciding what overdue work to run. The agenda is cheap; the gate is judgment — never delegate a gate.

Cadence checklists (each calls existing skills — don't re-implement)

• Daily (~15 min) — health check, NOT optimization. Via tracker: spend/CPC/clicks vs yesterday, budget pacing, disapprovals/policy, paused-by-error, sudden anomalies. Goal = spot what's BROKEN. No performance changes off one day of data.
• Weekly (1-2h) — via optimizer: search-term review (mandatory, both sources) — pull search_term_view (covers Search/Branded campaigns) AND PMax/Demand-Gen search category insights (campaign_search_term_insight; search_term_view does NOT contain PMax terms). Feed both to search_term_miner.py. Never propose blocking brand terms (brand intent = your own traffic, even at 0 conv). Then: budget pacing + reallocate to top performers, campaign tiering, automated-rules check. Apply the decision gates below.
• Monthly (2-4h): the full money-leak sweep — run D1-D14 in ${CLAUDE_PLUGIN_ROOT}/references/diagnostic-playbook.md: bid/target health (decode bidding_strategy_system_status; lower tROAS if LIMITED_BY_INVENTORY, scale if LIMITED_BY_BUDGET on-target), budget allocation (shift low-ROAS→high-ROAS), geo waste, dayparting, structure, Quality Score, PMax channel distribution (D8), ad copy/assets/extensions (D9). Any tROAS/budget move follows the Scaling Ladder (one variable, cooldowns). Plus conversion-tracking audit (measurement), attribution review. Pull a fresh audit if drift is suspected.
• Quarterly: strategic audit + plan alignment to business goals.