Dream Loop

7. Judge

Judging should ideally be done by a fresh subagent with a clean context each round (delegate_task), to keep it objective and cheap. Give the judge the latest screenshot, the concept, and (from round 2 on) the previous round's screenshot and verdict.

Mechanics: vision_analyze takes one image per call. Either have the judge make sequential calls (concept, then screenshot, then compare from memory of its own descriptions), or — better — stitch a labeled side-by-side composite with ImageMagick (convert concept.png shot.png +append compare.png) or PIL and analyze that single image.

Judge prompt:

You are an art director reviewing a real-time render against its concept art. Compare the screenshot to the concept and score it 0-10 using this ladder. The ladder is gated: a frame cannot score above a tier's cap until every requirement of the tiers below it is fully met. Be strict about the gates.

• Tier 1, shape (0-3): camera, framing, composition, and the position and rough scale of every major object match the concept. Layout, not finish: every major element present, in the right region of the frame (within ~10% of frame width/height), at roughly the right size (within ~25%). Right place and vaguely correct outline passes even if edges and surface are wrong; save precision nitpicks for Tier 4. Cap 3 until true.
• Tier 2, light and color (3-5): key light direction and color, overall exposure (no clipping to black or white), shadow depth, palette, contrast, atmosphere. Attend to reflections, glows, etc. Ensure the scene is not too bright or dark relative to the concept. Judge the whole frame, not tiny details (Tier 4). Cap 5 until lighting/reflections/color/ contrast are generally right.
• Tier 3, materials and surfaces (5-7): every surface reads as the right material at a glance: textures, roughness, translucency, wetness, reflections. Assets must not look procedural, blocky, smooth/plastic; frame-dominating elements should be properly sculpted and detailed. Cap 7 until true.
• Tier 4, fine detail (7-9): the small things. Nitpick relentlessly; inspect every little object up close. Layout aligns near-perfectly; materials extremely convincing. Cap 9 until right.
• Tier 5, indistinguishable (9-10): holds up side by side and zoomed in. Nitpick every pixel.

If a previous verdict and screenshot are provided: you are one reviewer in a sequence, not the first. Maintain consistency. First mark each previous directive LANDED, PARTIAL, or NOT DONE against the new screenshot; carry forward anything PARTIAL or NOT DONE. Don't reverse a prior directive unless the result is clearly worse — and if you do, say so and why.

Output format:

1. Score on the first line; "Tier N" (highest fully-passed gate) on the second. 1b. If given a previous verdict: the LANDED / PARTIAL / NOT DONE list.
2. "Blocking:" the specific failures of the next tier's gate. The builder must clear these before anything else counts. Name the element and the change, with magnitudes: "Rocks: replace the stacked ovoid boulders with one continuous fractured slab; cracks 2-5cm wide, dark interiors, add surface texture so they don't look flat/plastic" — not "the rocks look artificial".
3. Then at most 4 further directives from higher tiers, same style, ordered by points recoverable.

No non-actionable feedback ("this looks synthetic") — name the specific causes. Every directive must be actionable this round. Don't round up: if a gate isn't fully passed, the cap holds.

8. Exit criteria

• Score >= 8 and target FPS acceptable: done. Show the user the latest screenshot; ask if they want more iterations.
• Score >= 8 but FPS unacceptable: optimize — lossless wins first, then minimal-visual-impact ones. Re-judge afterwards to confirm no regression.
• Stall approaching (best score hasn't improved a full point in 2 rounds, or the judge named the same gap 3 times): stop incremental tweaks. Step back and ask what about the approach is capping the score. Make one big structural change in a round: swap asset strategy (sculpt in Blender, pull real models/textures/HDRIs if allowed), rewrite the lighting model, rebuild the composition, change the camera. Self-check carefully — big changes break things. Only repeat parameter tuning if you can articulate why it would work this time.
• Stalled (already tried a big structural change, score flat 3 rounds, judge is nitpicking or demanding intractable things like raytracing on a GPU-less machine): stop, tell the user why you're blocked, give options.
• Otherwise: address all or most heavy-hitting gaps this round, not just the top one — rounds are expensive. Prioritize gaps that move the needle most (lighting, textures, mesh detail). Only revert if the score dropped a full point or more; small dips are judge noise, and reverting a whole round throws out good changes with bad. If one change clearly regressed, undo just that change. Loop.

Follow-up loops

When building on an existing product (or the user re-invokes the skill for refinements), don't create new concept art in a vacuum — it may diverge from what exists. Instead capture a live screenshot of the current product and prompt image_generate to render the best possible version of it (current screenshot → AAA-graphics version of the same shot), then use that as the target. Multiple screens can run parallel judge loops if asked, at higher token cost.

Pitfalls

• Overbaked or oversimplified concept art (see failure modes) — fix the concept before building against it.
• Submitting half-baked screenshots to the judge; the self-check gate exists for a reason.
• Screenshot at a different resolution/aspect than the concept — unfair comparison, noisy verdicts.
• Tunnel-visioning on incremental tweaks when the judge says you're off base.
• Judging both images in one vision_analyze call — it takes one image; composite them first.
• Screenshotting before the scene loads/settles — add a wait after wait_for_load() for asset streaming and animation warm-up.

Verification

• .dream-loop/concept.png exists and passed the failure-mode review (and user confirmation, if generated).
• Each round has a screenshot, a logged self-assessment, and a judge verdict with score + tier + directives.
• Exit only via an explicit exit criterion; final screenshot shown to the user with the final score and FPS measurement.

Adapted from dream-loop by Anshu Chimala (MIT). Upstream license vendored as LICENSE.txt.