Video Brief — decide before you capture

Two failures make a video expensive, and neither can be fixed in the edit:

1. The footage does not exist. The shot needs a feature that is not built, a save state nobody has, or a build that has since moved on.
2. The platform rejects the footage. Apple states that an app preview must show only content within the app and must not film people interacting with a device — so an over-the-shoulder shot is rejected no matter how well it is cut, and no edit recovers it.

Both are decided before capture. That is what this skill is for. It does not edit, and it does not judge whether a finished cut is good — it produces the document that makes those later steps answerable.

Scope boundary. Whether the video worked is an analytics question, not a brief question. Do not put an unfalsifiable success claim in the brief; § 2 forces the alternative.

1. Resolve context before asking anything

1. Project type — the session hook prints PROJECT_TYPE. game routes to trailer/short defaults and game agents; web/mobile/service routes to demo/short and product agents. On unknown, ask before spawning anything (CLAUDE.md § Agent usage rules).
2. Read what the product already claims about itself — design/gdd/game-concept.md and design/gdd/game-pillars.md on a game track, product/prd/ and the product concept on a product track. A video that contradicts the pillars is a positioning problem, and surfacing it here is cheaper than after capture.
3. Kind — from --kind, else infer from what the user said and state the inference. The four kinds are trailer (store), demo (service/app), short (vertical feed), longform (devlog / tutorial).
4. Platform — from --platform, else ask. This is not cosmetic: it fixes duration, aspect ratio, whether sound exists, and several hard content rules.
5. Review mode — --review, else production/review-mode.txt, else lean.

2. Fix the job of the video

Use AskUserQuestion for the four fields in the template's § 1 that the project docs cannot answer: the specific viewer, where they meet it, the one belief they should leave with, and the next action.

Two things to hold the line on, because both are where briefs go soft:

• "Everyone" is not an audience and "awareness" is not a next action. Push once for a specific answer. If the honest answer really is awareness with no CTA, write that down — a stated null is a decision; an unstated one is a gap.
• The success metric must be falsifiable, or labelled as not attributable. Wishlist counts after a trailer drop are confounded by everything else shipping that week. Say so in the brief rather than promising attribution the team cannot produce.

Assume muted playback unless the placement guarantees otherwise. Store pages and feeds autoplay without sound, and that constrains the structure — it is not a subtitles task at the end.

3. Re-read the platform spec — every run, no exceptions

Fetch the vendor page for each target platform and record the date. docs/video-production-sources.md holds the links and the last known figures; it does not excuse you from re-reading them.

The reason is recorded in that file: while it was being written, two figures taken from third-party ASO blogs were wrong against the vendor documentation, and re-reading the vendor page is what caught them. Third-party summaries are not an acceptable source for a platform requirement.

The second row is BLOCKED, not CONCERNS, and the reason is the exit 3 rule in docs/deterministic-gates.md: a check that could not run has produced no verdict, and no verdict is not a pass. An unverified spec is an unknown spec, and § 6 already says capturing against an unknown spec is how footage gets thrown away. Writing the brief is still useful — write it, mark the platform section unverified, and do not call it READY.

Never present an unverified figure as a requirement.

4. Draft the structure, then the shot list

Work from docs/templates/video-brief.md. Two passes, and the order matters — the structure decides what footage is needed, not the reverse.

4.1 Structure

Fill the table for the resolved kind and delete the other three. The defaults have named origins (docs/video-production-sources.md); deviating is allowed, deviating silently is not — write the reason in the notes column.

For a game trailer, the two rules worth defending against pressure from inside the team:

• Genre legible by 0:10. Before the viewer can name what they are looking at, nothing else you show can land. A studio logo on a black screen at 0:00 spends the most valuable three seconds in the video.
• Every claim needs footage that proves it. "Tell, Show, Repeat" fails when a title card asserts something the following shot does not demonstrate. Walk the table and check each pair.

For a demo, the equivalent is: lead with the problem, and place the moment it clicks early. This one is weaker evidence than the trailer rules — it is aggregated practitioner advice, not a vendor rule or a named primary source (docs/video-production-sources.md flags the caveat). Present it as the default structure, not as a measured finding, and do not quote the retention percentages those posts carry.

4.2 Shot list

Every row names footage a person must capture. The column that earns the document is Exists?:

• A shot needing an unbuilt feature is a schedule dependency. Name it, and say what unblocks it. Do not quietly design around it — a trailer built only from what happens to exist today shows the game you have, not the one you are selling.
• A shot needing a scripted setup (a specific save, a staged encounter, a seeded account) is a capture-day task. It gets discovered at the worst moment if it is not written here.

Set the capture standard once — resolution, frame rate, HUD on/off, quality preset, no debug overlays, no placeholder art in frame — and apply it to every row. Mixed capture settings across shots are visible in the cut and cannot be repaired.

4.3 Who to spawn

Route by rules/route-hint.md. Route on this skill's artifact, not on the video's stakes — the artifact is a markdown brief, which is R1/R2 under rules/verify-route.md, and production route and verification route are independent axes that must not be inferred from each other. A store trailer being a public first impression argues for checking the brief harder, not for building it with more agents.

Tied between two routes? Take the lighter one.

Narrow tools at spawn: tools: Read, Glob, Grep. community-manager and art-director have Write by default, and this skill's premise is that the orchestrator is the sole writer. Read-only is guaranteed by permission, not by asking nicely (rules/subagent-collaboration.md § 3.1).

Write ownership, settled before spawning as § 2.1 requires:

Give each spawn full context; it cannot see this conversation.

5. Rights and content rules

Walk the template's § 4 and § 5 tables to the end, using only what docs/video-production-sources.md actually records for the target platform plus what you re-read in § 3. Do not state a store policy from memory — rules/claim-confidence.md § 2 puts regulatory and policy clauses in the verify-before-writing category, and platform policy is exactly that.

What is recorded, and therefore safe to assert: Apple requires previews to show only in-app content and tells publishers to display only material they hold the rights to; Google Play requires monetization off, which a copyright claim on borrowed music can defeat. Anything beyond those, verify or mark 추정.

An unresolved rights row makes the verdict BLOCKED, not CONCERNS. A brief that says "probably fine on the music" is how a rejection gets discovered after the capture budget is spent — and unlike a spec figure, this one is not merely unverified, it is unresolved.

6. Verdict

READY is about the brief, not the video. It says the expensive decisions have been made, not that the result will be good.

7. Write the brief

Ask before writing. Show the structure table and the shot list in conversation and confirm the path with the user before creating the file.

On approval, write to production/marketing/video-brief-<name>.md using docs/templates/video-brief.md. That path is the convention verify_policy.py P4 expects; do not invent a new one.

Next steps

• READY → capture against the shot list, then hand off to editing. The external tool this repo delegates to is named in docs/video-production-sources.md § "Execution tooling" — it makes a paid transcription call per source file, so run /api-cost-gate before the first invocation.
• BLOCKED on a shot dependency → the shot is a story. /create-stories, then revisit this brief when it lands.
• Store submission → /release-checklist and /launch-checklist own the submission itself; this brief is an input to them, not a substitute.
• A cut that deviates from the approved structure → back here for re-approval, not ratified after the fact.
• The same defect appearing in a second video → /lesson-log.

Error Recovery Protocol

If a spawned agent returns BLOCKED or errors:

1. Surface it immediately — "[AgentName]: BLOCKED — [reason]" before continuing to any dependent step.
2. Assess the dependency. The shot list depends on the structure; do not build one on a missing other.
3. Offer options via AskUserQuestion: skip and record the gap, retry narrower, or stop and resolve the blocker.
4. Always produce a partial brief. Never discard completed sections because one step blocked.