Compare Against the Reference
Give the agent a reference and tell it to keep comparing its output against it. The comparison is what closes the gap.
The prompt
"Keep comparing it against the reference and fix any differences." (Meng To)
Meng To built Three.js interfaces, Liquid Glass buttons and procedural models entirely in code this way, giving the model a video, URL or code reference and asking it to recreate it in one HTML file, then to verify a detail against the real object ("verify the result against the real MacBook"). Source: X/@MengTo, 2026-09-06
Why it works
A model judging its own output against a description will call it done too early, because the description is vague. A reference is not vague. Asking for repeated comparison makes each pass a diff: what in the reference is missing or wrong in the render. The agent has to look at its output, which many runs never do, and every fix has a target.
The same loop shows up wherever output is visual. Wes Bos's room scans check each modeled object against the photo taken from the same spot and angle. Source: X/@wesbos, 2026-10-22 Benji Taylor's Stageshow skill has the agent film its own UI animation and fix what the recording shows. Source: X/@benjitaylor, 2026-10-31
When to use it
- Recreating a design, product shot or scene from a screenshot, video or live site.
- 3D models and CAD parts, rendered from the reference's viewpoint.
- Any output a person will judge by eye.
The agent needs a way to see its output: a browser screenshot, a render, or a recording. Without one, "compare" is a guess.
Which skill runs it
No single skill. cad and the Text to CAD Skill Family family render and compare for 3D parts; agent-browser screenshots are how a web agent sees its own page.
At workflow scale
Asked to build an F1 showroom in Blender, Opus 5 under ultracode built a rough car, then ran about 50 agents for 10 hours: 27 built detailed parts, each iterating against its own render; 5 built competing monocoque versions judged head to head, and the winner was refined; 18 inspectors zoomed into every region hunting defects. The post names no outside reference: each agent compared against its own render, and the inspectors' defect hunt stood in for one. Source: X/@aipulseda1ly, 2026-07-25
img2threejs packages the loop: it rebuilds the object in one reference photo as procedural Three.js code, gated by a render-versus-reference comparison, so the result is editable 3D instead of a static asset. Source: X/@MengTo, 2026-07-20; Source: hoainho/img2threejs
Timeline
- 2026-11-03 | Page created from the three.js recreation prompts, with the room-scan and Stageshow loops as the same pattern. Source: X bookmarks, 2026-09-06 to 2026-10-31