Visual director gate

Visual Director Gate

Use this skill after the story is clear and before final render approval. It does not enforce one fixed style. It enforces a stronger visual communication standard for each new video.

Required Inputs

Prefer these artifacts:

• current episode.json;
• first-frame or cover still;
• scene contact sheet;
• caption-midpoint contact sheet;
• raw key frames for important UI actions;
• previous published episode first-frame and scene sheet, when available.

If artifacts are missing, stop and generate the cheapest missing artifact first. Do not approve final render from memory.

Review Order

1. Read references/account-visual-brief.md.
2. Run the score script on available artifacts:

python3 skills/visual-director-gate/scripts/score_video_artifacts.py \
  --episode-json /path/to/episode.json \
  --first-frame /path/to/first-frame.png \
  --scene-sheet /path/to/scene-sheet.png \
  --caption-sheet /path/to/caption-sheet.png \
  --previous-scene-sheet /path/to/previous-scene-sheet.png \
  --out /path/to/visual-gate-report.md

For a first episode with no baseline, add --first-episode. Otherwise the previous scene sheet is mandatory.

3. Inspect the actual images. Fill or amend the report with visual scores.
4. If the score is below the gate threshold, revise the visual direction before rendering final.

Scorecard

Use 100 points. Passing threshold is 82. A video below 70 is a redesign, not a polish pass.

• First-frame memory: 20
  • 0-8: topic visible but no memorable claim or asset;
  • 9-15: claim is readable but visual identity is generic;
  • 16-20: one-second silent frame clearly communicates topic, creator asset, and why it matters.
• Information hierarchy: 20
  • 0-8: screenshots or captions compete without a dominant reading order;
  • 9-15: main action is understandable after listening;
  • 16-20: the eye knows what to read or click first in every scene midpoint.
• UI focus and magnification: 15
  • 0-5: key UI is merely present;
  • 6-11: key UI is enlarged but not always centered or timed;
  • 12-15: each critical click/copy/input target becomes the unique focal point, normally at least 20% of frame width or with an unmistakable pointer.
• Motion purpose: 15
  • 0-5: movement is decorative;
  • 6-11: movement creates energy but not always attention;
  • 12-15: push, pan, cursor, wipe, or reveal changes the viewer's understanding.
• Caption discipline: 10
  • 0-3: captions explain what the picture failed to show;
  • 4-7: captions are short but sometimes carry the main action alone;
  • 8-10: captions label and reinforce what is already visually obvious.
• Evidence and privacy: 10
  • 0-3: fake UI, unclear proof, or privacy risk;
  • 4-7: real assets are present but proof/privacy is uneven;
  • 8-10: real artifacts are visible, sensitive data is masked, and evidence does not distract.
• Progress over prior episode: 10
  • 0-3: weaker or merely different;
  • 4-7: comparable, with one clear improvement;
  • 8-10: visibly more precise, memorable, or mature than the chosen baseline.

Blocking Rules

Block final approval when any item is true:

• first frame cannot be understood silently within one second;
• title, narration hook, and first visual imply different promises;
• a critical UI click/copy/input target is not visually findable within two seconds;
• captions are the only place the viewer can understand the main action;
• text overlaps, clips, or looks smaller than the scene's purpose allows;
• a privacy-sensitive token, cookie, account identifier, or personal field is exposed;
• final scene sheet is visually weaker than the approved proxy or previous published baseline without a deliberate reason.

Design Bias

Prefer:

• large claims and step labels paired with real UI evidence;
• one focal target per shot;
• local zoom, crop, cursor, and micro-labels over broad boxes;
• motion that carries attention from old state to new state;
• fewer visual objects with stronger hierarchy;
• visible improvement over the prior episode rather than strict sameness.

Avoid:

• screenshot collage without a dominant message;
• decorative movement that does not clarify anything;
• tiny UI shown only for authenticity;
• overusing callout boxes after the user has rejected broad selection frames;
• explaining boundaries that are not part of the core viewer task.

Output

Every visual gate pass should produce a short report with:

• score by category;
• pass/fail decision;
• top three visual weaknesses;
• required edits before final, if any;
• evidence paths inspected.