We asked Claude Opus 5.5 for a high-detail technical teardown of the original PS5 disc edition, twice, in two empty folders. The only difference was whether Skillshed was switched on.
The prompt. An exploded view of the CFI-1000's main components, the assembled console one switch away, every part labelled, internals based on public teardowns and anything unconfirmed marked as illustrative, delivered as a page that opens from disk with SVG and PNG downloads. It never mentions skills.
The two runs. Both headless, claude -p, with only the folder's own settings loaded so nothing from our machine leaked in. The plain run had no plugins at all. The Skillshed run had the defaults plus the interfaces pack, the way skillshed use interfaces sets it up.
What we measured. Wall time, cost, tokens per model, turns, tool calls and which skills fired, read from each run's transcript. Both ran at the same time on the same Mac, with Node 22 and Claude Code 2.1.308. The transcripts and the runner are in the repo.
What this isn't. A benchmark. Each side ran once, so the gap could move on another day. Our benchmarks run every task three times per arm with a scored rubric; a drawing has no rubric, so here you judge the result yourself.