Candidate APolished default
Clear, but interchangeable after the logo is removed.
Taste engine for AI builders
Make taste
prove itself.
Compare real references, rebuild in isolation, and release only what survives the evidence.
Sample analysisLocal evidence only
Candidate BEvidence-led
Wins on product truth, hierarchy, and recall.
One loop. No taste theater.
The system searches widely, lands carefully, and records why the winner deserved to survive.
- 01Read the product
Audience, task, truth, constraints.
- 02Acquire taste
Relevant mechanisms, plus anti-references.
- 03Run the tournament
Distinct candidates, blinded pairwise choice.
- 04Repair the winner
One defect class per reversible patch.
- 05Prove release
Browser evidence and weakest-link gates.
A skill, not a sentence.
The website compiles context into a versioned Taste IR. Your local Agent receives rules, references, authority, and a stopping condition it can inspect.
Brief + evidenceTaste IRSkill / MCP / CI
taste-ir.jsonpinned and portable
{
"product_truth": "release requires evidence",
"strategy": "greenfield",
"taste": {
"prefer": ["clear hierarchy", "owned mechanism"],
"reject": ["template mimicry", "score theater"]
},
"authority": {
"change_level": 5,
"production_branch": false
},
"stop_when": "release_bar_met"
}Same engine. Different authority.
Maturity changes the size of the patch, not the standard of evidence.
GreenfieldLevel 5
Find the direction.
Generate broad visual worlds, then implement one winner.
RescueLevel 2-3
Fix the shared cause.
Preserve product structure. Repair components and tokens.
GovernedLevel 0-1
Prevent the drift.
Audit, localize, and propose a small reversible patch.
Release Bar
A binary decision with visible evidence.
No parent average can hide a broken mobile flow, keyboard trap, or false claim.
Candidate stateBLOCKED
Missing mobile error recovery
Visualpass
Interactionpass
Responsiveblocked
Accessibilitypass
Functionalpass
Product validation targets
Commercial proof, without the fake beauty claim.
These are pilot targets, not current performance and not a promise that 95% of sites become top-tier design.
No severe visual or interaction defect
Eligible projects reach Release Bar
Blind preference for the Taste Engine version
Reduction in matched human edit time