Proof, not promises.
Softproof runs each AI design skill on the same briefs as the same model with no skill at all, checks every render with code, has a blind jury score it, and prices it. You see what the skill actually adds.
See the leaderboardBrowse single-shot renders
Same brief (card-06), same model. Over all six briefs the blind jury scored taste-skill 7.76 and the baseline 7.22.
Illustrative pair selected where the baseline and selected condition differ on technical acceptance. The full gallery retains every single-shot output. The Claude juror rates Claude-generated outputs, blinding and equal weighting do not remove family bias, and no human validation has been performed. Scores combine fit to brief, visual quality and brand fit.
On these six briefs our instruction pack passed 8/12 at $0.0026457 per accepted asset, the best technical pass rate and lowest generation accounting cost per acceptance in the single-shot track. It also had the lowest composite jury mean, 7.17/10, versus baseline 7.22 and taste-skill 7.76. This does not establish general visual or full-workflow superiority.
Same model. Same briefs. One skill each.
Single shot: the skill's instructions, one model call, no tools, no revisions, on Claude Haiku 5.5. The common configured output setting was 10,000 tokens. One ranked call reported 16,552 aggregate output tokens and one loop call 15,063, mostly thinking tokens. This setting is not a verified hard aggregate cap. The shaded row is the model with no skill. A render counts as accepted only when it passes every code check, and the jury's scores are kept separate from that. Six briefs with one generation each, rendered at two correlated sizes, is a small sample, so read it as what happened on these briefs, not a general ranking. Blind jury: claude-sonnet-5-5, gpt-6.1-sol, gemini-4-argon-medium. The Claude juror rates Claude-generated outputs. Blinding and equal weighting do not remove family bias. No human validation has been performed. The composite averages fit to brief, visual quality and brand fit, rather than visual quality alone.
On these six briefs our instruction pack passed 8/12 at $0.0026457 per accepted asset, the best technical pass rate and lowest generation accounting cost per acceptance in the single-shot track. It also had the lowest composite jury mean, 7.17/10, versus baseline 7.22 and taste-skill 7.76. This does not establish general visual or full-workflow superiority.
| # | Skill | Passes every check | Blind jury | Cost per accepted asset | Time per brief | By check |
|---|---|---|---|---|---|---|
| 1 | taste-skilljury's pick Leonxlnx/taste-skill f624f37 |
50.0% 6/12 assets |
7.76 | $0.0039985 | 30.1s | |
| 2 | ui-ux-pro-max | 33.3% 4/12 assets |
7.51 | $0.0057029 | 29.2s | |
| 3 | frontend-design anthropics/skills dbd4588 |
41.7% 5/12 assets |
7.44 | $0.0058729 | 41.8s | |
| 4 | naksha-studio Adityaraj0421/naksha-studio f527525 |
58.3% 7/12 assets |
7.24 | $0.0034401 | 26.6s | |
| 5 | baseline no skill loaded |
33.3% 4/12 assets |
7.22 | $0.0051436 | 30.9s | |
| 6 | impeccable pbakaus/impeccable 2f46b7a |
25.0% 3/12 assets |
7.19 | $0.0077446 | 31.8s | |
| 7 | designermost checks passedcheapest accepted render adityareddy-dev/softproof 96086a6 |
66.7% 8/12 assets |
7.17 | $0.0026457 | 28.5s |
Bars, left to right: contrast, overflow, safe zones, type scale, palette, copy. Cost is generation-only CLI list-price accounting on existing subscriptions, including failed outputs, divided by the renders that passed. Jury costs are excluded. These values are not actual billing. Time per brief is generation model-call time, excluding installation and rendering. Full numbers are in summary.json, with receipts and raw juror answers in the results archive.
Our designer, with its loop.
Our designer also runs with its render, check and revise loop on the same briefs. That gets extra rounds the single-shot table doesn't allow, so it isn't ranked against the others and has no jury score. Skills built around their own loops get an agentic track next round.
| Run | Passes every check | Generations per brief | Cost per accepted asset | Time per brief |
|---|---|---|---|---|
| designer-loop | 100.0% 12/12 assets | 1.3 | $0.0025845 | 41.3s |
All ranked outputs retained.
Every single-shot render in this round is retained, including failures. The gallery shows 1200x627 with links to the 1080x1350 renders. Open a brief to compare instruction packs side by side. See complete results and check receipts. The gallery excludes the unranked loop. Its complete evidence includes final renders, every intermediate round, failed checks and usage receipts. Total loop generation accounting cost was $0.03101377 versus $0.0211659 for our single shot, so a lower cost per acceptance does not mean lower total spend.
card-01 Less noise. Clearer signals. · tideline brand · linkedin
A product announcement with a clear headline and a simple signal diagram.
card-02 Made slowly. Used daily. · ember brand · linkedin
An editorial collection card, using abstract vessel geometry rather than photos.
card-03 Your next mile starts here. · volt brand · linkedin
An energetic training launch with a large numeric anchor.
card-04 Build a calmer on-call week. · tideline brand · linkedin
An event announcement. Date, time and CTA must be easy to find.
card-05 Good things earn their place. · ember brand · linkedin
A quote-like editorial card with strong typographic rhythm and a small issue marker.
How a round works.
Fixed briefs
Copy, audience, sizes and a DESIGN.md brand file. Copy is fixed, so writing isn't judged.
Pinned skills
Installed from their own repos at a commit, never copied here. Same model and configured setting within the single-shot track.
Rendered
Headless Chrome at every size. Scores come from pixels and the DOM, not the model's own account.
Checked
Contrast, overflow, benchmark-policy safe zones, type scale, palette and copy, all in plain code. Safe zones do not guarantee platform crops.
Judged, priced
Three blind jurors from different model families, plus tokens, time and cost per accepted render.
Jurors see shuffled, unlabelled renders and the brief, never skill names or check results. Results get published whoever comes out ahead, including our own designer.
Pinned commits, model IDs and briefs let you rerun the procedure. Outputs are stochastic, the CLI has no supported seed, and identical outputs are not guaranteed. The harness limits matter: filters, blend modes, occlusion, complex masks and path-based lettering need visual review. Raster images, canvas and iframes fail pending separate inspection. Technical acceptance does not establish visual quality.