Skill demos 5
One made-up product, Tallbook, built 5 times. Each version came from a fresh agent allowed one design skill, or none at all. Same brief, same words, so only the design changes.
Read the briefNo skill
No skill loaded
What Claude builds by default, with no design skill loaded. The control.
$1.08 at API prices21.7k tokens written34 tool calls4.5 min
Design taste
design-taste-frontend
Deals a researched art direction from a library of gallery sites, then holds a strict no-ornament house style.
$4.60 at API prices77.3k tokens written80 tool calls22.1 min
Impeccable
impeccable
Works like a design director: picks a mode and a point of view, then holds a craft floor.
$2.55 at API prices48.4k tokens written50 tool calls10 min
Apple design
apple-design
Apple’s approach for the web: physical motion and springs, materials and depth, careful type.
$2.01 at API prices53.2k tokens written25 tool calls9.5 min
Emil design eng
emil-design-eng
Emil Kowalski’s polish: component feel, easing and timing, the details you only notice when missing.
$1.72 at API prices40k tokens written35 tool calls7.3 min
What each build used
Counted from each builder agent's own transcript and priced at the API list prices for claude-opus-5-5: $4.00 per million input tokens, $5.00 to write the cache, $0.20 to read it and $20.00 per million output tokens. A Claude plan doesn't bill per token, so this is what the same work would cost through the API. It leaves out the work of setting the agents up and wiring the pages in.
| Version | Cost | Written | Read fresh | Read from cache | Requests | Tool calls | Time (all five at once) |
|---|---|---|---|---|---|---|---|
| No skill | $1.08 | 21,720 | 81,433 | 1,212,346 | 20 | 34 | 4.5 min |
| Design taste | $4.60 | 77,306 | 247,673 | 9,065,216 | 61 | 80 | 22.1 min |
| Impeccable | $2.55 | 48,435 | 160,416 | 3,919,270 | 35 | 50 | 10 min |
| Apple design | $2.01 | 53,242 | 119,269 | 1,751,160 | 20 | 25 | 9.5 min |
| Emil design eng | $1.72 | 39,998 | 117,624 | 1,637,768 | 20 | 35 | 7.3 min |
| All five | $11.96 | 240,701 | 726,415 | 17,585,760 | 156 | 224 |
Written is output, thinking included. Read fresh is input the model saw for the first time, mostly the skill's instructions and the files it opened. Read from cache is the same context re-read on every step, which is why it is large but cheap.
No build was capped or hurried: each ran its skill's whole process and stopped when its own checks passed, so a skill with more steps costs more. Time is first step to last, with all five running at once on one machine and sharing one dev server and browser, and it includes waiting on screenshots and searches. Read cost, tokens and tool calls for the fair comparison, and time as a rough guide.