Zebrafish UGC Ad Pack

Follow the phases in order. Do not skip a phase or a gate. The zebrafish router skill's Setup and Gates (spend, claim, identity, publish) apply before and during this playbook.

Preconditions (MUST)#

• Router setup is complete: context script ran, campaign/product knowledge loaded, and ../zebrafish/references/tool-belt.md read (hosted: https://zebrafish.dev/skills/zebrafish/references/tool-belt.md).

• Required inputs, all present:

  • [ ]

    Product brief or URL.

  • [ ]

    Product image/reference.

  • [ ]

    Target customer.

  • [ ]

    Offer/CTA.

  • [ ]

    Claims that are allowed and claims to avoid.

  • [ ]

    Variant count.

  • [ ]

    Platform and aspect ratio.

• If a required input is missing, STOP and ask for it. Do not guess. Claim boundaries especially: never infer the allowed-claims list.

Phase 1 - Intake and Brief (gate: brief confirmed)#

Restate the brief back to the user before any planning work:

• Outputs: variant count, platform(s), aspect ratio(s), and target duration per variant.

• Target customer, offer, and CTA.

• Allowed claims and forbidden claims, verbatim.

• Presenter/voice expectations and identity constraints. Approval needs.

STOP until the user confirms scope. Do not plan or spend against an unconfirmed brief.

Phase 2 - Plan (gate: spend approved)#

1. Extract product facts and allowed claims into PRODUCT_CLAIM_SHEET. Every scripted claim must trace to this sheet per the router claim gate.

2. Build the HOOK_MATRIX across angles: pain, curiosity, proof, offer, comparison, and demo. Map each variant to one hook angle.

3. Write the VARIANT_SCRIPT_PACK, one script per variant.

4. Set PRESENTER_STYLE_RULES: product visual style and presenter/voice constraints. No real-person likeness or voice without the router identity gate cleared.

5. Write the SHOT_LIST and ASSET_PLAN.

6. End with a production plan: assets per variant, which models (chosen via list_models), unit counts, and estimated cost.

STOP for spend approval per the router spend gate before any paid generation. Approval covers the stated variant batch only; extra variants, resolutions, or asset types re-ask.

Phase 3 - Produce#

1. Generate product visuals, clips, and audio through approved runtime paths only.

2. Assemble each variant and label it by hook angle from HOOK_MATRIX.

3. Score every variant in SCORECARD: hook clarity, product focus, pacing, brand fit, and claim safety.

4. Poll every generation and run to a terminal state. As each output completes, record its provenance fields (provider, model, prompt, generation/run IDs, references, cost) into the delivery document. Draft the delivery document during this phase, not after.

Phase 4 - Verify#

Apply the video gates from ../zebrafish/references/verification.md (hosted: https://zebrafish.dev/skills/zebrafish/references/verification.md):

• Claim safety: every spoken, captioned, or on-screen claim in every variant traces to PRODUCT_CLAIM_SHEET; forbidden claims are absent.

• Hook labeling: every variant carries the correct hook-angle label and its opening seconds actually deliver that hook.

• Inspect actual frames per variant: hook, product moments, CTA.

• Check aspect ratio, duration, and variant count against the brief.

• Confirm every URL is reachable and cost evidence is captured.

Run node ../zebrafish/scripts/verify.mjs --strict <delivery file> (the router skill ships the same script; agents may use either copy). A non-zero exit fails the gate: fix findings or move them to ## Unverified with a reason, then re-run until it exits clean.

Phase 5 - Deliver#

MUST use the delivery template from ../zebrafish/references/delivery.md (hosted: https://zebrafish.dev/skills/zebrafish/references/delivery.md): a # Delivery: title, then ## Outputs, ## Provenance, ## Verification, ## Unverified, in that order. URLs first: each variant URL labeled by hook angle.

After the template sections, append the skill's extra artifacts: hook matrix summary, script pack summary, and scorecard summary. Update the PROVENANCE and VERIFY_REPORT artifacts to match.

Studio Artifacts#

Create or maintain PRODUCT_CLAIM_SHEET, HOOK_MATRIX, VARIANT_SCRIPT_PACK, PRESENTER_STYLE_RULES, SHOT_LIST, ASSET_PLAN, SCORECARD, PROVENANCE, and VERIFY_REPORT.

Provider Policy#

• Use approved Zebrafish/FAL runtime provider paths for product visuals, clips, and audio.

• Treat external UGC/ad generators as research evidence only unless approved in Zebrafish runtime.

• Do not use real-person likeness, voice, or persona workflows without explicit consent, retention, and deletion rules.

Stop Conditions#

STOP and ask before:

• spending on generation without approval;

• using real-person likeness, voice, or identity references;

• creating claims not present in the source material;

• using an unapproved external provider.