Google Ads — Plan

Turn a budget + business context into a concrete plan and an honest forecast. Read everything from the context and the audit/measurement outputs — no hardcoding, no guessing where data exists.

STEP 0 — Load context + GATE

1. Read account-context.yaml (budget, AOV, margin_tiers, brand_terms, vertical, guardrails) and the latest audit + measurement outputs from the working directory. If the context is missing, run setup first — never plan spend without it.
2. Measurement gate: if measurement = FAIL, STOP — do not plan spend on broken tracking. Tell the user to fix tracking first. WARN is allowed but carried into the plan as a risk note.
3. Confirm campaign_defaults.daily_budget (or ask). Output is written to the working directory.

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

• Scout (haiku) — the forecaster.py run; reading existing campaign budgets.
• Routine (sonnet) — STEP 1 mode-detection pull (existing campaigns + spend), STEP 4 deriving CPC/CVR from account data. Dispatch as general-purpose sub-agents; return raw, don't decide the mode.
• Judge (main session) — STEP 0 measurement gate, the campaign mix + sequencing, budget split, forecast interpretation (band not promise), ramp roadmap, risks. The data is cheap; the plan is judgment.

STEP 1 — Detect mode: LAUNCH vs EXPANSION

Read existing ENABLED campaigns + their spend/performance. The plan differs sharply:

• LAUNCH (no/low history) — start simple, lean on benchmarks, conservative ranges, one or two campaign types, no target ROAS during learning.
• EXPANSION (established account) — read actual scale and find gaps. Read the real budget scale, don't assume small (e.g. a mature account may run $1k+/day across many campaigns). Recommend additions/reallocations against what already exists, not a from-scratch structure. State which mode you detected and why.

STEP 2 — Campaign mix (which types, in what order)

Use the vertical priority in ${CLAUDE_PLUGIN_ROOT}/references/vertical-defaults.md + what's missing in the account:

• ecommerce: PMax/Shopping → Branded Search → Search → Demand Gen.
• leadgen / saas / b2b: (Branded) Search → Search → Demand Gen. Justify each: goal it serves (acquisition / defense / scale), and why now. Coordinate Branded Search + PMax brand exclusion so they don't cannibalize (honor pmax-brand-exclusion guardrail).
• Non-brand Search is a HARVEST/CONTROL layer on PMax discovery, not a from-scratch keyword bet. Recommend it once PMax has discovered converting non-brand queries worth controlling (high volume/value/margin, or needing distinct messaging — check the dual-source search-term data: search_term_view + campaign_search_term_insight per PMax). If there's nothing proven to harvest yet, hold Search and let PMax keep discovering. When you do recommend it, hand builder-search the harvest list, not a category list.

STEP 3 — Budget split + learning minimums

• Split the daily budget across chosen types by goal and expected efficiency.
• Learning minimum: each campaign / asset group / ad group needs ~15-30 conversions/month to exit learning. With a limited budget, run FEWER campaigns well rather than starving many. Compute the max number of campaigns/AGs the budget can actually feed and say so.
• Apply margin_tiers: high-margin lines can run at lower ROAS; reflect that in the split.

STEP 4 — Forecast (derive, don't guess) — see ${CLAUDE_PLUGIN_ROOT}/references/forecasting-and-benchmarks.md

• EXPANSION: derive CPC and CVR from the account's own recent data; project clicks = budget / CPC, conversions = clicks × CVR, revenue = conversions × AOV, ROAS = revenue / spend, CPA = spend / conversions. Present as a band, not a point estimate.
• LAUNCH: use vertical benchmark bands; present WIDE ranges and label them estimates with stated assumptions. Never present a benchmark forecast as a promise.
• Always show the inputs (CPC, CVR, AOV, budget) with their source. If AOV is not verified from store data (context connections.store = missing), say so and label the revenue/ROAS line UNVERIFIED — connect store for a real AOV; use a clearly-flagged provisional only with the user's OK. Never present a guessed AOV as fact — it silently distorts every revenue/ROAS number downstream.
• Runnable helper: python ${CLAUDE_PLUGIN_ROOT}/skills/plan/scripts/forecaster.py --budget B --cpc C --cvr R --aov A [--target-roas T] → prints conv/revenue/ROAS/CPA bands + learning capacity (how many campaigns the budget can feed).

STEP 5 — Ramp roadmap

Phase the bidding per campaign_defaults.bidding_ramp (e.g. Maximize Conversion Value with no target during learning → Target ROAS once volume matures). Respect ramp discipline: don't set a target before enough conversions; raise tROAS ≤0.3x per step with a wait between steps; honor change-event-cooldown.

STEP 6 — Success criteria + risks

• Define what "working" means per campaign (target CPA/ROAS by phase, learning-exit timeline).
• List risks (measurement WARNs, thin budget vs learning minimum, seasonality).

Output — GOOGLE-ADS-PLAN.md (to the working dir)

• Mode (launch/expansion) + rationale.
• Campaign mix + sequencing.
• Budget split table + max-campaigns-the-budget-can-feed.
• Forecast band with inputs shown.
• Ramp roadmap + success criteria + risks.
• Hand off: for each chosen campaign type, point to the matching builder-* skill.

To build / refine later

• Runnable forecaster (scripts/forecaster.py). Done.