/ads

Claude Ads prompt patterns

These original examples specify routing, worker contracts, uncertainty, and mutation safety. Use them when behavior is subtle; do not load them for simple calculations or a clearly routed platform question.

1. Route by intended artifact

User: “Use this product photo to make four Meta ad concepts.”

• Route to ads-create when the user needs concepts, messaging, hooks, or briefs.
• Route to ads-generate only when a validated brief exists and image files are requested.
• Route to ads-photoshoot when the primary request is product-photography transformation rather than composed advertising concepts.
• If the request includes both concept and rendering, run create, obtain approval, then generate. Do not skip the intermediate manifest.

2. Bounded platform worker

{
  "objective": "Audit Google conversion measurement for the supplied window",
  "scope": ["conversion taxonomy", "tag and enhanced-conversion evidence", "deduplication"],
  "exclusions": ["creative", "budget", "account mutation"],
  "evidence_policy": ["account export", "official current sources"],
  "privacy_class": "confidential",
  "mutation_authority": "read-only",
  "inputs": ["run:20260711-demo/account-snapshot.json"],
  "output_contract": "Finding[] v1.0.0",
  "verification": ["schema validate", "cite every failed control"],
  "recovery": ["return needs_input for missing tag evidence"]
}

The worker returns findings and recovery hints to the conductor. It does not write google-audit-results.md, calculate a platform score, or broaden into budget work.

3. Missing evidence

User: “My Meta CPA doubled. Pause the worst ads.”

If the supplied data lacks conversion lag, attribution definition, spend, sample size, recent changes, or current remote state:

1. Record the observed CPA change only if windows are comparable.
2. Return needs_input for a mutation.
3. Offer a read-only diagnostic and an explicit data request.
4. Do not apply a fixed CPA multiple or infer that the worst observed ad caused the account-level change.

4. Approved mutation

An approval is valid only for the exact mutation plan presented:

{
  "account_id": "redacted-demo",
  "object_id": "campaign-123",
  "operation": "pause",
  "before": {"status": "ACTIVE"},
  "after": {"status": "PAUSED"},
  "reason": "Operator-approved policy containment",
  "blast_radius": "one campaign",
  "idempotency_key": "sha256:...",
  "verification_window": "immediate plus 15 minutes",
  "rollback": {"operation": "restore_status", "value": "ACTIVE"}
}

If remote state no longer matches before, stop and regenerate the plan. Approval for one campaign, budget, or date never transfers to another.

5. Research and independent verification

The research worker proposes source records and affected claims. A separate source verifier checks authority, date, scope, geography, methodology, license, and claim-to-source fit. Only then may the conductor update canonical references and tests. A community post can create a research lead; it cannot establish an official API or policy claim.

6. Partial full audit

If Google, Meta, and tracking succeed but the requested Amazon worker fails authentication:

• Preserve the successful platform results.
• Mark the bundle partial.
• Exclude Amazon from numerical portfolio health and display its missing weight.
• Return the authentication recovery step.
• Never title the deliverable “complete multi-platform audit.”