She reads the chart. She makes one cut. Nothing vital is ever touched.
The smallest cut that heals.
You know the other guy — long ponytail, oval glasses, replaces your fifty lines with one. Great. But sometimes the line he deletes was the one checking your input, and an independent 480-build benchmark caught him trimming everyday bad-input handling on 5 of 24 tasks, and his ladder re-deliberates every single turn of an agent session, quietly billing you for the meditation.
Scalpel puts a surgeon in your agent instead. She plans the operation once, reuses what the body already has (your codebase, the stdlib, the platform) before grafting anything new, makes the smallest incision that actually heals — and never severs an artery. Validation, error handling, security, and accessibility are anatomy, not fat.
You ask for a date picker. Your agent installs flatpickr, writes a wrapper component, adds a stylesheet, and starts a discussion about timezones.
With scalpel:
<!-- scalpel: the browser has one -->
<input type="date">Same one-liner the lazy guy writes — but when you ask for an email validator, scalpel's version still rejects garbage input. Minimal never means flimsy.
Before cutting, the surgeon reads the chart, then prefers what already exists:
1. Does this need to exist? → no: don't operate (YAGNI)
2. Already in this codebase? → grep for it, reuse it
3. Stdlib / native platform? → use it
4. Installed dependency? → use it — never add a new one for a few lines
5. Only then: new code → the minimum that works
Then three rules the ladder guys don't have:
- One incision. The plan is decided once, then executed — no re-deliberating the operation every response. That per-turn meditation is where minimalism skills quietly spend your tokens.
- Anatomy list. Input validation at trust boundaries, error handling, security, accessibility: never cut, and the final diff is checked for exactly that.
- Silent hands. Code first, at most two lines of prose.
[code] → skipped: [X], add when [Y].
Measured on the same harness ponytail uses (vendored from their repo, MIT — same 5 tasks, same execution-based correctness gate, same LOC counter), four arms, 3 Claude models × 10 repeats per cell (n=50), no cache.
Scalpel vs ponytail, head-to-head:
| model | LOC | tokens | cost | latency | correctness |
|---|---|---|---|---|---|
| Haiku 4.5 | −12% | −23% | −10% | −4% | 100% vs 96% |
| Sonnet 4.6 | −13% | −25% | −12% | tie | tie (100%) |
| Opus 4.8 | −30% | −27% | −20% | −13% | tie (100%) |
Beats or ties on every cell, loses none — and vs the no-skill baseline: ~90% less code at ~4× lower cost. Ponytail's only correctness failures in the run were real: it prescribed email_validator, a package that wasn't installed, and the code died — the exact "artery" scalpel's dependency rule guards. Full tables, failure notes, and honesty caveats: benchmarks/results/2026-07-03-single-shot.md.
Agentic (real headless Claude Code sessions on a real FastAPI+React repo, ponytail's own harness, 12 features + 7 adversarial safety tasks): a statistical tie with ponytail — scalpel edges LOC, time, and over-engineering flags; ponytail edges cost by ~3%; both 100% correct, 100% safe, both ~69% less code than no-skill, and the completeness judge confirms scalpel's low LOC isn't stub-shipping (3.00/3, min 3, every cell). Where ponytail's README admits its per-task cut is "near zero on already-minimal code", scalpel holds the same floor while winning the single-shot suite outright. Details: benchmarks/results/2026-07-03-agentic.md.
Reproduce:
cd benchmarks
ANTHROPIC_API_KEY=... npx promptfoo@latest eval -c promptfooconfig.yaml --repeat 10
npx promptfoo@latest viewClaude Code (plugin):
/plugin install scalpelOr drop skills/scalpel/SKILL.md into .claude/skills/scalpel/ in any project, or paste it into the system prompt of any agent that takes one.
Scalpel governs what you build, not who you are. "stop scalpel" / "normal mode" turns it off. If you explicitly ask for the fuller version of something, she builds it — a surgeon doesn't argue with informed consent.
Benchmark harness vendored from DietrichGebert/ponytail (MIT) so every comparison is apples-to-apples on their scoreboard. Caveman arm from JuliusBrussee/caveman (MIT).
MIT