{"id":9275,"plugin_id":"plugin_asdk_app_6a1ea93307208191b5fe82494caf73a8","kind":"skill","collection_source":null,"comparison_source":null,"observed_at":"2026-09-30T22:54:33.901Z","digest":"0ac8f746ee9c2d50035662a8b24ed1403f6ba99544ffba0dbe805f9d3fb66e4f","against":null,"payload":{"name":"creative-testing","description":"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.","included_files":[{"relative_path":"agents/openai.yaml","size_in_bytes":535},{"relative_path":"references/experiment-design.md","size_in_bytes":1829}],"skill_md_contents":"---\nname: creative-testing\ndescription: 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.\n---\n\n# Creative Testing\n\nCreate tests that isolate a meaningful decision and produce learning the team can reuse—not a pile of unrelated variants.\n\n## Guardrails\n\n- Do not claim statistical significance from ordinary organic social comparisons, small samples, unequal delivery, or platform-reported totals without a valid experiment design.\n- Hold audience, offer, placement, timing, and CTA constant when testing a creative variable unless one of those is the declared variable.\n- Preserve factual claims and required disclaimers across variants. Never make a test “stronger” by inventing proof or urgency.\n- Image or video generation spends credits. Confirm ambiguous generation and avoid generating variants that do not test a defined hypothesis.\n- Create variants as drafts and require explicit approval before scheduling or queueing.\n- Do not automatically declare a winner from the highest raw engagement count; match the decision metric to the objective.\n\n## Workflow\n\n1. 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.\n2. When a baseline exists, call `social_getSocialMediaAccounts`, then retrieve relevant aggregated, range, and post analytics. Distinguish observed patterns from hypotheses.\n3. Write one falsifiable hypothesis: changing **X** for **Y audience/context** should improve **Z metric** because **reason**.\n4. 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.\n5. Define the control and two to four purposeful variants. Each variant must express a distinct strategic alternative, not superficial synonym changes.\n6. 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.\n7. Produce a test matrix with invariant elements, variable, hypothesis, assets, account/placement, run window, minimum practical evidence, and decision rule.\n8. 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.\n9. Create each execution with `social_createSocialMediaPost` and `action: \"draft\"`, using required platform settings. Never publish one variant early and call it a fair comparison.\n10. 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.\n\n## Experimentation Standard\n\n- Prioritize high-leverage uncertainty. Test the promise or proof before button color, emoji, or trivial copy edits.\n- Separate exploration from validation. Early tests can identify promising territories; later tests should isolate and confirm the driver.\n- Build variants from different audience tensions or persuasion mechanisms, not random creativity.\n- Evaluate platform delivery effects, audience overlap, spend, timing, and sample imbalance before attributing performance to creative.\n- Stop tests that create brand, legal, reputational, or customer-experience risk regardless of short-term metrics.\n- Turn each result into a reusable rule with scope: what worked, for whom, where, under what conditions, and what remains unknown.\n\n## Output\n\nLead 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.\n"},"changes":[],"summary":"First saved snapshot. No earlier version is available for comparison.","summary_kind":"deterministic","summary_metadata":{}}