# Creative Testing

> Design, produce, and evaluate disciplined social creative tests for marketers. Use when the user asks for A/B tests, hook tests, creative variants, message experiments, format tests, offer or CTA tests, an experimentation roadmap, ways to improve a campaign systematically, or draft variants whose results can produce a reusable marketing learning.

- Skill: `celeryhq/creative-testing` (Agent Skill, multi-file: 3 files)
- Install (CLI): `npx skillmds@latest add celeryhq/creative-testing`
- Raw SKILL.md: https://api.skillmd.com/api/skills/celeryhq/creative-testing/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Product & Planning
- Author: celeryhq (https://skillmd.com/u/celeryhq)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/celeryhq/creative-testing

---


# Creative Testing

Create tests that isolate a meaningful decision and produce learning the team can reuse—not a pile of unrelated variants.

## Guardrails

- Do not claim statistical significance from ordinary organic social comparisons, small samples, unequal delivery, or platform-reported totals without a valid experiment design.
- Hold audience, offer, placement, timing, and CTA constant when testing a creative variable unless one of those is the declared variable.
- Preserve factual claims and required disclaimers across variants. Never make a test “stronger” by inventing proof or urgency.
- Image or video generation spends credits. Confirm ambiguous generation and avoid generating variants that do not test a defined hypothesis.
- Create variants as drafts and require explicit approval before scheduling or queueing.
- Do not automatically declare a winner from the highest raw engagement count; match the decision metric to the objective.

## Workflow

1. Define the business decision: what choice will change if this test succeeds? Resolve objective, audience, offer, channel/placement, conversion path, current control, constraints, and available volume.
2. When a baseline exists, call `social_getSocialMediaAccounts`, then retrieve relevant aggregated, range, and post analytics. Distinguish observed patterns from hypotheses.
3. Write one falsifiable hypothesis: changing **X** for **Y audience/context** should improve **Z metric** because **reason**.
4. Select one primary variable: hook, promise framing, proof type, visual treatment, opening frame, format, CTA language, creator/brand voice, or offer framing. Use [references/experiment-design.md](references/experiment-design.md) to control confounds.
5. Define the control and two to four purposeful variants. Each variant must express a distinct strategic alternative, not superficial synonym changes.
6. Choose a primary decision metric and guardrails before production. Examples: qualified reach/video hold for attention, saves or substantive engagement for utility, clicks/leads/bookings for response, and negative feedback for audience cost.
7. Produce a test matrix with invariant elements, variable, hypothesis, assets, account/placement, run window, minimum practical evidence, and decision rule.
8. If new creative is authorized, use `$generate-image` or `$generate-video` with reusable asset storage. Keep composition, product, and brand constants unless visual treatment is the tested variable.
9. Create each execution with `social_createSocialMediaPost` and `action: "draft"`, using required platform settings. Never publish one variant early and call it a fair comparison.
10. After the run, use `$social-performance-analyst` to compare results. Record result, confidence/limitations, learning, next decision, and follow-up test. Retain a control until a challenger wins under a credible comparison.

## Experimentation Standard

- Prioritize high-leverage uncertainty. Test the promise or proof before button color, emoji, or trivial copy edits.
- Separate exploration from validation. Early tests can identify promising territories; later tests should isolate and confirm the driver.
- Build variants from different audience tensions or persuasion mechanisms, not random creativity.
- Evaluate platform delivery effects, audience overlap, spend, timing, and sample imbalance before attributing performance to creative.
- Stop tests that create brand, legal, reputational, or customer-experience risk regardless of short-term metrics.
- Turn each result into a reusable rule with scope: what worked, for whom, where, under what conditions, and what remains unknown.

## Output

Lead with the decision and hypothesis. Then show the controlled matrix, draft/asset status, measurement and stopping rules, validity risks, and the learning record the team should complete after results arrive.