Ad creative
Use this method to produce paid ad copy that survives platform constraints and creates a clear next test. It produces drafts and recommendations; it does not publish ads, change budgets, or claim performance that the evidence does not support.
Context and intake
Use supplied brand guidance, approved claims, offer details, landing-page copy, and performance data when available. Ask only for missing details that affect copy quality or compliance:
- platform and format;
- offer, audience, and awareness stage;
- value proposition and evidence;
- approved voice, prohibited claims, and regulatory constraints;
- performance data when iterating.
If the user needs a fast draft, make the narrowest practical assumptions and label them.
Generate from scratch
1. Define three to five distinct creative angles. 2. Vary concept, hook, headline structure, visual direction, and proof—not only synonyms. 3. Generate platform-ready variants for each angle. 4. Validate character limits against `./references/platform-specs.md`. 5. Recommend a small test and the signal that would count as learning.
Iterate from performance data
1. Identify top and bottom performers using the metric supplied by the user. 2. Separate creative signals from audience, offer, landing-page, and tracking signals that the data cannot distinguish. 3. Identify winning structures and weak dimensions without calling a result proven from a thin sample. 4. Extend one or two supported patterns and add one genuinely different angle. 5. Record the hypothesis, change, metric, result status, and next review point.
Visual or video direction
Read `./references/generative-tools.md` for a provider-neutral visual brief and quality review. Use `creative.visual_brief` from `./references/host-operations.md` only when the user asks for an asset and the host has authorized it. Otherwise return the brief and intended placements.
Output format
```markdown ## Assumptions - Platform: - Offer: - Audience:
## Angles 1. [Angle] — [why it could work]
## Creative ### Angle: [Name] #### Headlines ([limit]) 1. "[copy]" ([count])
#### Primary text or descriptions ([limit]) 1. "[copy]" ([count])
## Recommended test - [variants, metric, and decision rule]
## Performance summary - Evidence considered: - Signal and limitation:
## New variations - [variation and what it changes]
## Creative diversity - Score: [0-10] - Weakest dimension: - What changed:
## Iteration log - Hypothesis: - Change: - Metric and review point: - Proposed learning record: ```
Offer CSV when the user asks for ten or more variations.
Quality and safety
- Show character counts and a trimmed alternative for over-limit copy. - Keep responsive-search-ad headlines independent. - Match the landing-page promise and use only supportable claims. - Avoid fake urgency, unsupported superlatives, guarantees, and unapproved regulated-industry claims. - For algorithmic-feed creative, assess concept, format, visual, hook, and headline diversity. Score the set from 0 to 10. Below 8, rebuild the weakest dimension before adding variants; do not pad the set with near-duplicates. - When performance evidence is supplied, use `learning_record.propose` from `./references/host-operations.md` only if authorized. Otherwise return the proposed learning record in the iteration log. - Return a proposed launch or optimization plan; publishing and spending remain explicit host-authorized actions.