video-request-architect

Video Request Architect

Use When

• The scene logic, storyboard, beat sheet, or prompt plan is already approved.
• You need a provider-agnostic request architecture before mapping to a video model.
• A long script must be split into self-contained Seedance-ready segments.
• The target creative format or aspect ratio must stay explicit through request mapping.

Do Not Use When

• Use benchmark-to-brief first when the campaign pattern or hook evidence is still unclear.
• Use reference-decode or reference-contract-builder first when reference meaning or copy boundaries are unresolved.
• Use prompt-preflight-qa when the request is already drafted and needs QA before generation.

Core Rule

Prompt architecture comes before provider syntax.

Write the visible clip in time order before choosing model fields. Every final request must be self-contained; do not rely on a model remembering previous segments, earlier prompts, or unstated reference bindings.

This architect skill builds provider-agnostic request structure. It must not call providers, submit jobs, upload media, or poll results.

Lock the PostPlus creative format before provider mapping. Default to short_form_vertical / 9:16; use instagram_meta_ads / 3:4 when the brief is for Instagram Meta Ads. The selected ratio must appear in the architecture, segment contract, and final render request.

Default Workflow

1. Lock the experiment goal: hook replication, camera realism, persona continuity, benefit proof, CTA readability, or another narrow test.
2. Build the scene before fields: first frame, visible action, spoken-line support, payoff moment, product timing, and binding references.
3. Convert the scene into provider-agnostic blocks: role, goal, hook logic, reference policy, output spec, look/tone, timecoded beat sheet, camera, sound intent, and product policy.
4. If the target duration exceeds 15 seconds for Seedance, require a timecoded beatSheet with ordered, non-overlapping startSeconds and endSeconds. Split into independently generatable segments before provider mapping.
5. Print or return the request skeleton before the final provider request: goal, segment type, hook logic, viewer question, reference policy, duration, creative format, aspect ratio, camera grammar, product policy, and main risks.
6. Map to provider fields only after the architecture is coherent: model, duration, aspect ratio, media references, quality, and resolution.

Output Shape

The artifact contains a JSON architecture with:

• creativeFormat, targetAspectRatio, and duration
• segmentType, goal, hookLogic, and viewerQuestion
• referencePolicy, cameraGrammar, productPolicy, and mainRisks
• skeleton for the request checkpoint
• segmentContract with Seedance segmentation status and per-segment handoff data when segmentation is required

Stop Conditions

• Stop when required user intent, source evidence, or owned input artifacts are missing and guessing would change the result.
• If an owned CLI or script command fails, report the exact error and stop. Do not bypass the failure with metadata-only answers, readiness probing, local payload rewrites, fallback providers, or unpublished tools.

Handoff

• Hand request-architecture.json to the provider mapping or generation skill.
• For hook or segment uncertainty, hand back to benchmark-to-brief or the human script review gate.
• For reference uncertainty, hand to reference-decode or reference-contract-builder.

Public Command Boundary

• Choose the smallest matching command or workflow from the user input and run it directly.
• If an owned CLI or script command fails, report the exact error and stop. Do not bypass the failure with metadata-only answers, readiness probing, local payload rewrites, fallback providers, or unpublished tools.
• This public skill is instruction-driven. Produce the artifact described by the workflow directly from the available evidence.
• Do not call private provider/runtime paths or unpublished local tools.
• If the CLI returns a quote-confirmation challenge, run postplus quote confirm --json --challenge-file <challenge.json> and retry with the returned token.