Review award caliber art direction

Review Award-Caliber Art Direction

Hold the line

Judge the current rendered artifact, never the intention, effort, source-code sophistication, test count, or previous score.

Treat 10/10 as exceptional and rare. Award it only when the category is simultaneously:

• authored and specific;
• coherent with the whole product;
• instantly legible at the target size;
• polished through motion and state changes;
• free of unresolved high-impact defects;
• strong enough to survive comparison with award-winning shipped work.

Issue PASS only when every required category is honestly 10/10. Otherwise issue STOP. Do not average strengths over weaknesses, round up, or pass because the product improved.

Choose the operating mode

Infer the mode from the request:

• Review-only: inspect and report. Do not edit files.
• Improve-and-review: inspect, implement the highest-leverage fixes, refresh evidence, and repeat.
• Release gate: inspect the launch candidate and return only PASS or STOP with decisive evidence.

If the user names a target rubric, preserve it exactly. Otherwise use the ten-category scorecard in references/scorecard.md.

Load the required references

Read all three references before starting a complete review:

1. references/evidence-protocol.md for capture coverage and proof integrity.
2. references/scorecard.md for category rubrics, score caps, and premium-quality thresholds.
3. references/report-template.md for the verdict and re-review structure.

Run the review

1. Establish the visual contract

Extract the target platform, viewport, genre, primary action, emotional promise, visual identity, accessibility constraints, and explicit quality bar from current files and the user request.

Write a one-sentence contract:

The product must make [player action] feel [emotion] through [distinctive visual language] at [target size/platform].

Use that contract to distinguish purposeful restraint from unfinished work.

2. Audit the real current state

Inspect the running artifact when available. Use static assets or source only to diagnose what the rendered evidence reveals.

Record the exact build identifier, URL or file, viewport, device scale, state, time or frame, motion setting, and capture timestamp. Cache-bust after material changes. Never reuse a pre-fix screenshot as current evidence.

Keep evidence classes explicit:

• Observed: directly visible or measurable in the current artifact.
• Inferred: a conclusion supported by multiple observations.
• Unverified: not inspected or not reproducible.

Never convert unverified evidence into a pass.

3. Capture a complete evidence matrix

Capture the whole product, not one hero frame. Include:

• first impression and onboarding;
• ordinary play at low, medium, and high visual pressure;
• every signature character or enemy state;
• anticipation, action, impact, recovery, and idle motion;
• success, failure, reward, progression, pause, and result UI;
• shortest supported phone and representative target phone;
• reduced-motion state when supported;
• 120 px downsample and grayscale/value proofs;
• audio/visual onset and impact alignment when sound is in scope.

Follow the exact matrix and browser-resource rules in the evidence protocol.

4. Diagnose systems, not isolated blemishes

Trace each defect to the visual system that produced it. Prefer root-cause labels such as:

• decorative layer detached from world physics;
• inconsistent projection or light source;
• generic UI component language;
• silhouette family collapse;
• animation without anticipation or recovery;
• procedural geometry competing with authored material;
• reward louder or more legible than danger;
• pasted asset with no contact, occlusion, or received light;
• AI-like texture abundance without a governing shape grammar.

Do not prescribe more particles, glow, gradients, detail, or camera shake without naming the perceptual problem they solve.

5. Score independently

Score each category from its own current evidence. Apply every automatic cap in the scorecard. A category may score lower after improvement if fresh evidence exposes a deeper inconsistency.

For every score below 10, provide:

• current-build evidence;
• why it prevents 10/10;
• the smallest high-leverage system fix;
• the exact state or capture needed to verify the fix.

6. Prioritize by perceptual leverage

Rank defects:

• P0: breaks gameplay comprehension, identity, physical continuity, or basic trust.
• P1: immediately prevents premium or award-level perception.
• P2: visible polish debt that compounds across repeated play.
• P3: local refinement with low systemic impact.

Fix P0 and P1 systems before micro-polish. One coherent material rule is worth more than ten ornamental additions.

7. Iterate without grading your own homework

When implementation is authorized:

1. Fix the highest-leverage defect at its source.
2. Refresh the build identifier or cache key.
3. Recapture the affected state and at least one neighboring state.
4. Recheck 120 px, grayscale, motion continuity, and UI collision.
5. Run relevant regression checks.
6. Re-score from zero using only fresh evidence.

Prefer an independent reviewer for the final gate when available. Give the reviewer the current artifact and rubric, not the intended answer or prior score.

Protect resource discipline

Use at most one browser tab for the review unless the user explicitly permits more. Reuse it across states and close it immediately after capture. Confirm the final open-tab count is zero.

Do not leave local servers, capture processes, temporary windows, or generated evidence running after the review unless the user asked for monitoring.

Refuse common false passes

Return STOP when any of these remain:

• the best screenshot hides a weaker ordinary-play state;
• a static frame looks good but motion ghosts, teleports, or loses mass;
• grayscale collapses player, threat, objective, or depth planes;
• the product is readable only above its real mobile size;
• UI looks imported from a different product or generic web template;
• characters are detailed but share one generic silhouette or face grammar;
• sacred geometry behaves as a logo overlay rather than a physical world rule;
• visual richness comes from noise rather than hierarchy;
• audio is clipped, inaudible, generic, or late relative to contact;
• evidence is stale, partial, inferred, or missing.

End with brutal self-critique

Answer all seven questions before a final verdict:

1. What still feels generic?
2. What still feels boring?
3. What still feels forgettable?
4. What still feels amateur?
5. What still prevents million-download screenshot appeal?
6. What still prevents award consideration?
7. What still prevents a 4.8+ player rating?

Do not soften the answers to protect prior work.

Example invocations

• Use $review-award-caliber-art-direction on this mobile game and refuse a pass until all ten areas are 10/10.
• Review the enemy animation and UI/world integration at 390x844. Give fresh evidence only.
• Run an award-caliber release gate on the current build. Inspect temporal states, 120 px, grayscale, and AV sync.