The four exhibits

Each of the four exhibits is a frozen, self-contained build of the same website: same brief, same starting code, same Astro 7 / Tailwind 4 stack, same four page types (home, blog index, blog post, about), same content. Only the design tool differed: Claude's own frontend-design skill, Impeccable, Taste, and gstack each took the identical starting point and produced its own result. Every exhibit is marked noindex, so it may not be crawlable by search engines; the stats and context on this page are the canonical record of what each tool did. Open all four and compare them directly. That is what they are for.

Arm 1: what did frontend-design do?

frontend-design is Claude's own design skill. Its process asks nothing: it self-fills every gap in the brief, runs one visual self-review, and commits. It was the cheapest arm by roughly six times the cost of the next cheapest, and it is the real benchmark of the four: a complete, shipped, reviewed design for a modelled $1.12. It paired Archivo for display type with Faustina for body text.

Type

Archivo (display) + Faustina (body)

Intake questions

0

Tool calls

71

Wall clock

~26 minutes

Tokens

~169,000

Review rounds

1 self-review

External model

No

Modelled cost

$1.12

Commit

91835f5

Arm 1's model is inferred from strong same-session evidence, not a recorded field.

View the arm 1 exhibit →

Arm 2: what did Impeccable do?

Impeccable is an internal adversarial system: a builder, a reviewer sitting behind an explicit fix|ship gate, and a documenter that re-derives its output from source rather than summarizing the builder's claims. It ran two full rounds before it would pass itself, at roughly six times arm 1's token cost. Its two-round gate caught a hard-banned kicker it had shipped itself, an h3 heading set quieter than its own body text, and a quarter of its own mark system left uncoded. It chose the identical type pair as arm 1 with no contact between the two runs.

Type

Archivo (display) + Faustina (body)

Intake questions

21 (15 from the sheet, 6 "your call")

Tool calls (design)

258

Wall clock (design)

~54 minutes

Tokens (design)

1,064,198

Review rounds

2 (fix|ship gate)

External model

No

Modelled cost (design)

$7.02

Commit

43587c5

Arm 2 deployment

Tool calls

15

Wall clock

2.5 minutes

Tokens

48,971

Cost

$0.32

Impeccable's image-generation step was skipped because it bills an external key the brief never authorised. This is a partial exercise of the tool.

View the arm 2 exhibit →

Arm 3: what did Taste do?

Taste (design-taste-frontend) runs a templated six-slot intake, one self-review pass, and applies the strictest copy rules of the four arms. It costs the most to port to production because it hands over React that must be translated to the site's Astro/Tailwind stack.

Type

Archivo (display), Literata (body), Geist Mono (third)

Intake questions

18 (9 from the sheet, 6 "your call")

Tool calls (design)

116

Wall clock (design)

~48 minutes

Tokens (design)

284,287

Review rounds

1 self-review

External model

No

Modelled cost (design)

$1.88

Commit

5b3b250

Arm 3 deployment

Tool calls

40

Wall clock

7.1 minutes

Tokens

68,648

Cost

$0.45

View the arm 3 exhibit →

Arm 4: what did gstack do?

gstack ran design-consultation, then design-review: a structured 13-finding review with an external model brought in as a second voice, one commit per finding. It caught the single most valuable defect of the whole run: a rule named .stamp--alert was structurally unable to win against the surrounding context rules (specificity 0,2,0 against the modifier's 0,1,0), and the three files meant to consume it never applied it, so the one place this design uses a second hue would never have rendered red anywhere a reader meets it. No build gate, screenshot, or visual review would have caught that fault. Of its intake, all eight "your call" answers fell on axes the brief deliberately left open, and it named two conflicts against the stated vision and rejected both.

Type

Archivo (display), Piazzolla (body), Martian Mono (third)

Intake questions

23 (12 from the sheet, 8 "your call")

Tool calls (design)

151

Wall clock (design)

~46 minutes

Tokens (design)

409,632

Review rounds

13-finding review

External model

Yes (2 codex exec calls)

Modelled cost (design)

$2.70

Commit

b580b68

Arm 4 deployment

Tool calls

75

Wall clock

38.6 minutes: a Cloudflare control-plane success the data plane did not follow; the DNS record did not appear until the binding was deleted and re-attached after a backoff

Tokens

100,079

Cost

$0.66

gstack invokes codex exec as a built-in step, so an OpenAI model contributed findings and a design direction to this arm. It is a multi-model ensemble where every other arm was single-model: "model = Opus 5" cannot be claimed for arm 4, even though its subdomain, claude-gstack-v1, says otherwise on its face. Separately, gstack's AI-mockup variant board never ran because its design binary is not installed on this machine, so this is also a partial exercise of the tool.

View the arm 4 exhibit →

What should a reader discount before comparing these four?

Two of the four arms, Impeccable and gstack, never ran their flagship step: Impeccable's image generation and gstack's AI-mockup variant board were both skipped for reasons outside the tools' own design logic. No comparison across the four exhibits should treat either arm as a full exercise of its tool. Separately, all four exhibits still ship an internal ticket reference ("Bio pending", pointing at TASK-005) on their About pages, because the brief's copy was frozen and no tool would invent a bio to fill the gap. The exhibits were deployed faithfully as each tool produced them, rather than patched afterward to remove that line.

See how the brief and starting conditions were set in the methodology, or the full cross-arm comparison in the results.