Before-After Ad

Purpose

Create approval-ready before-and-after ad concepts and shot scripts. Make the starting state inspectable, show a plausible product-use behavior, and resolve with controlled visual evidence rather than an unsupported transformation claim.

Produce a Before-After Grammar Plan, shot script, one approval checkpoint, and direct downstream handoff after approval.

Workflow

1. Ground The Brief

Read the smallest useful source context first:

• product facts, approved claims, usage instructions, and substantiation
• audience, persona, concern, desired visible result, and duration
• product, creator, scene, audio, and style references
• prior reference-video analysis or benchmark notes when relevant

If the user supplies raw reference videos without an analysis, use media-analysis first. Do not infer style from filenames, thumbnails, or a single frame.

Lock these before writing:

• starting_state
• desired_result_state
• pressure_carrier
• product_capability
• approved_proof
• claim_boundary
• persona_and_concern_map
• product_appearance
• duration_and_format
• reference_duties

Ask only when a missing fact would change the product, proof, claim, identity, or workflow. Do not invent treatment timelines, dosage, medical causality, guaranteed outcomes, or product appearance.

2. Design The Proof Grammar

Read references/style-grammar.md.

Produce a Before-After Grammar Plan with these fields:

• core_human_state
• pressure_carrier
• baseline_evidence
• comparison_control
• product_entrance
• product_gesture
• result_evidence
• camera_language
• framing_rules
• movement_rules
• edit_rhythm
• sound_mode
• text_mode
• claim_boundary
• continuity_locks
• generator_risks

Choose one primary proof route:

• direct comparison: baseline → use → matched result → optional split view
• discovery: external trigger → inspection → use → result
• routine montage: concern → repeated normal use → result portrait

Do not copy a reference creator, exact scene, wording, music, brand marks, or recognizable personal details. Transfer only reusable camera, movement, edit, proof, sound, and text behavior.

3. Write The Shot Script

Read references/shot-script-rubric.md.

Return a shot table with exactly these columns:

time | visual | framing/composition | camera motion | action | sound | subtitle/text | proof duty | purpose

After the table, return Production Locks with:

• persona_lock
• concern_map_lock
• product_lock
• scene_lock
• before_after_match_lock
• audio_lock
• claim_lock
• post_only_elements

Keep each shot physically executable. Use one clear action per shot. Treat captions, arrows, circles, split-screen assembly, end cards, disclaimers, and comparison labels as post-production elements rather than generated scene content.

After the script, ask whether the direction is approved. Stop before media generation unless the user approves or explicitly asks to continue into production in the same request.

4. Handoff After Approval

Read references/downstream-handoff.md and hand off directly:

• Use image-batch-runner first when persona, product lifestyle, concern-map, matched before/after, first-frame, or per-clip scene images are needed.
• Use video-batch-runner when the approved script and required hosted assets already exist and the user wants renders. Pass the approved plan, script, locks, facts, claim boundary, source basis, and reference duties without a redundant brief or second confirmation.

Let the downstream skill own model selection, request schema, upload rules, render execution, polling, and manifests.

Quality Bar

• Make the problem inspectable before a normal product-use gesture.
• Match proof-critical before and after framing, posture, light, and concern area.
• Put the emotional shift after evidence and preserve natural texture.
• Use one restrained text system, hard cuts, and subject-driven motion.
• Separate observed state, testimony, and substantiated product facts.

Stop Conditions

• Stop when required product appearance lacks an approved reference or creation permission.
• Stop when the requested result or timeline exceeds the approved claim boundary.
• Stop when comparison cannot be honest or references contain no transferable grammar.