Ask an agent to review an overdesigned component and it does something strange: it writes a genuinely competent critique — and then buries it. In the baseline this skill was built against, the reviewer produced twenty-plus findings across five frameworks, graded them with traffic-light emoji, mixed additions in with removals, and shipped an entire rewritten component nobody asked for. Thorough, and useless as a decision aid.
Deter is the counter-shape. A review is five parts and nothing else: the surface’s job in one sentence, removals each priced as a cost to that job, at least one thing kept with its reason, at most one missing piece of function, and the smallest next step. Frequency is the multiplier throughout — a flourish on a daily surface costs its price every day.
The rules with teeth:
- Findings, not files. It will not write a “revised version” unless you ask after reading the review. An unrequested rewrite makes the decision for you and hides the reasoning inside diff noise.
- A review that keeps nothing has not understood the design — it has only disliked it. The Kept section is mandatory.
- Taste needs a mechanism. “Gradients feel dated” is not a finding; “the gradient band pushes the renewal date below the fold on a card checked daily” is.
- No severity theater. No icons, no scores — ranking is prose order, and the reason is the finding.
Where a project keeps a Canon record, Deter checks removals against it first — a decision the team already made and recorded outranks the skill’s defaults, cited by entry.
The name works twice — it deters excess, and it’s a hat-tip one vowel away — and the phrase is rendered correctly — less, but better, not “less is more,” which belongs to Mies. Verified against its baseline: same overdesigned billing card, and the skill produced eight priced removals, one kept, one missing, zero files written.
Canon’s own, MIT, installable from the Canon marketplace.