skills/ai-host-video/references/creative-review.md

# Creative motion review

Creative quality is judged by an LLM against the episode. Local operations preserve exact request payloads and measure media; they do not
decide editorial quality.

## Advisory editorial review

The author reviews the plan while writing it. An independent reviewer is optional,
only when actually available and authorized or requested by the user. This MCP tool
surface supplies no reviewer tool. Do not invent a tool call. Use the relevant script,
coverage, direction, plan and reference constraints for your own comparison.

`motion-review.md` is optional working notes. No PASS token, waiver form, reviewer
availability or report format is required for request preparation. A missing or empty
review never stops production. Resolve actual defects against the brief; do not invent
an independent approval. If a reviewer is unavailable, continue the author review in
the same run. Honor a user's instruction to skip this advisory review directly.

The author reviews the decision chain, not isolated symptoms:

1. **Attention ownership** — each beat uses ON-CAMERA, VISUAL, or PLAYBACK for the
   semantic reason in `edit-plan.md`. Asset availability, implementation convenience,
   or missing negative space did not choose the role. A screen-filling composition
   under host narration appears only when the best integrated host-present composition
   would be less readable or editorially weaker for a concrete reason.
2. **Editorial value** — every treatment supports attention, readability, rhythm,
   emphasis, continuity, understanding, or another episode-specific viewing need
   with a clear connection to the script and intended viewer experience.
3. **Motion purpose** — each treatment performs its stated action. Removing competing
   objects, pointing to evidence and transforming a retained word must be visible;
   showing the corresponding labels is not equivalent.
4. **Episode progression** — each appearance of a recurring subject or component fits
   its current beat and the episode's established visual language.
5. **Contrast and restraint** — quiet passages have a positive performance or
   attention reason; active passages earn their density. The review does not fill gaps
   or reward event count.
6. **Episode coherence** — art direction, components, and animation behavior form one
   visual language while each treatment fits its content, placement, readability,
   ownership, and intended experience.
7. **Feasibility and fidelity** — required information states, anchors, planned-camera
   behavior, fallbacks, and readable holds are feasible from the submitted plan,
   reference footage, and available media without pre-solving JSX.

Challenge zero-asset plans against the strongest rejected image/video. A breed
countdown must let viewers inspect breed appearance; rank/name cards alone cannot.
Implementation convenience and "the host can explain it" are not user constraints.
Name a stronger beat-specific treatment when information is lost. Repeated cards,
even a larger last card, do not establish progression. Correct the underlying
editorial decision, without blanket layout bans. Counts and durations are
diagnostic facts, not thresholds; they never determine the verdict.

Apply useful findings to the owning artifact. Do not start a repeated review loop or
restart paid generation to settle a disagreement. The actual rendered master receives
the final visual/audio review below; a plan review cannot substitute for that evidence.

## Final rendered-motion review

Review actual encoded-master frames and available audiovisual playback as described
in [delivery](delivery.md). Compare them yourself against the script, selected motion
direction, edit plan and required states. The asynchronous video-analysis tool accepts
only its documented media id or YouTube URL; it has no creative prompt/script argument
and its scene summary is not an independent frame-level creative verdict. Clearly
report any unavailable playback or audio review. Within the observed evidence:

- compare the actual before/after states and the action between them with the plan;
  identify missing, weakened or mistimed behavior and evaluate any declared fallback;
- judge style fidelity through composition, motion character, readable states and
  transition grammar, not palette alone. Compare preset work with its selected
  local style reference and episode adaptation; compare auto/custom with the authored
  direction. Verify each use against readability, placement, and ownership;
- judge still duration by its inspection or emphasis value. Keep a useful quiet hold;
  restore missing semantic development instead of adding movement merely to fill time;
- reject motion that adds activity without supporting an approved editorial purpose;
- confirm that host presence, supporting media, visually quiet passages, and
  transitions produce a coherent episode-level rhythm; restraint should feel composed,
  and successive states should develop the idea rather than repeat an entrance. Check
  whether a background change restarts the host window or briefly exposes full-screen
  host, and whether word impact lands on the intended word rather than phrase start;
- name a stronger beat-specific alternative for every creative failure rather than
  asking generically for “more motion.”

If the rendered failure is an implementation weakness, revise layout, components,
timing, easing, transitions, or other `edit.jsx` mechanics and repeat the affected
proof states without changing the editorial plan. Revise the editorial plan only if the fix changes
treatment intent, ownership mode, selected media, required semantic states or their
order, speech anchors, fallback, or the episode's editorial motion direction. Any
reviewer opinion is evidence; the author still resolves contradictions
against the user's brief and the approved source artifacts.

Resolve observed findings in the edit and inspect the changed output. Keep useful
notes in any convenient form; no review file, PASS token or schema is required.
Judge intentionally dark passages in context through actual video inspection.