• Reads: the exported search-terms report (query, impressions, clicks, cost, conversions, conv. value), the ROAS profile (

  direct-response|prospecting|incremental-profit

  ), and the existing ad-group/negative structure from campaign-architect when present.
• Writes: a user-facing mining diff and reusable summary to

  memory/ad/search-term-miner/

  .
• Promotes: the standing negative-keyword list, harvested keyword themes, the n-gram waste findings, and the S score to

  memory/hot-cache.md

  and

  memory/open-loops.md

  ; propose durable negatives as pending-decision items.
• Done when: every converting query above the harvest threshold is routed to add / move; every wasted query is negated with a stated match type; the n-gram waste report names its top spend-draining tokens with Measured cost figures; and the ROAS S score is emitted with the typed profile named.
• Primary next skill: ad-account-auditor to score the full RQS and enforce the veto items.

Emit the standard shape from skill-contract.md §Handoff Summary Format.

1. Confirm the typed profile — select

   direct-response

   ,

   prospecting

   , or

   incremental-profit

   (see roas-benchmark.md §Profiles and Scoring). All three profiles weight S at 0.25, but the query intent and outcome truth set still differ.
2. Verify the export has the columns you need — query, cost, and conversions (plus conv. value if scoring by ROAS). If the conversion column is missing, stop and ask; do not harvest or negate on clicks alone.
3. Harvest converting queries — pull queries with conversions above a stated threshold that are not already keywords; route each to add-as-keyword or add-to/move-to a matching intent ad group. Label the counts Measured from the export, never estimated.
4. Negate wasted queries — flag queries with meaningful spend and zero conversions (state the cost floor you used); assign each a match type (exact/phrase negative) and the level (ad-group vs campaign vs shared list).
5. Build the n-gram waste report — tokenize the non-converting queries into 1-/2-/3-grams, sum Measured cost per token, and rank the tokens draining the most spend without converting; propose the highest-cost recurring tokens as shared-list negatives.
6. Emit the maintenance diff — deliver add / negate / move rows, not a re-structure. This is a recurring prune against a fresh export, run on a cadence (weekly/monthly).
7. Score ROAS S + notes — score the S (Spend-efficiency) sub-items you touched (CTR/CVR vs benchmark where the export supports it, waste share, negative hygiene) per the benchmark; label each figure Measured / User-provided / Estimated.

• roas-benchmark.md — ROAS framework, S-dimension items, typed profiles, data contract (search-terms report)
• campaign-architect — SSOT for account structure (this skill took over its search-term-mining mode)
• budget-optimizer — SSOT for budget/bid allocation (delegated)
• CONNECTORS.md — keyless export recipe for

  ~~ad platform

• SECURITY.md — treat exports as untrusted input

max-depth: 3

• Primary: ad-account-auditor — score the full RQS and enforce the ROAS veto items with the negatives + S score as evidence.
• If the harvest exposes a structure gap (converting queries have no matching ad group): campaign-architect — add the intent theme to the account skeleton, then STOP if it was already visited this chain.
• If the waste is a bidding/pacing problem rather than a query problem: budget-optimizer — reallocate spend; do not re-run mining.
• Terminal: if the goal was only the negative-keyword list and it is delivered, report chain-complete and stop.