A skill for coding agents · 171 words · MIT
Cut the code,
not the product.
Minimal-code skills save money by shipping less of what you asked for. Whittle tells your agent to ship all of it, and to spend nothing on the rest.
Same model, same brief, a hidden checklist
Minimal-code skills report how little got written. Nobody reports what got left out, so we measured it.
Build a personal kanban board as a web app. I want a complete product I could use every day, not a demo.
-
No skill
91–94%delivered
$0.30–0.46 per run
The whole product, at full price.
-
Ponytail 5
82–83%delivered
$0.19–0.20 per run
The cheapest. Its own skill says a vague request gets "the smallest version that does the core job".
-
Whittle
94–95%delivered
$0.26–0.27 per run
Keeps the product and costs about 30% less than no skill on average.
Claude Sonnet 5.5 in Claude Code, two apps from scratch (kanban, notes), 76 checks, two rounds of 5 runs per cell. Cost is list price. Every run and patch: 5 October, 14 October.
Which features got built
Runs out of five in which the kanban board had the feature. A full row of squares means it was built every time. Measured 14 October against Ponytail 5. These are the six features it left out of most of its boards.
| Feature | No skill | Ponytail 5 | Whittle |
|---|---|---|---|
| Archive or trash | 5 of 5 | 0 of 5 | 5 of 5 |
| Checklists or subtasks | 5 of 5 | 0 of 5 | 5 of 5 |
| Work-in-progress limits | 5 of 5 | 1 of 5 | 3 of 5 |
| Reordering columns | 3 of 5 | 2 of 5 | 2 of 5 |
| Versioned saved data | 1 of 5 | 0 of 5 | 5 of 5 |
| Motion that respects reduced motion | 0 of 5 | 0 of 5 | 5 of 5 |
The whole skill
Three paragraphs. No hooks, no modes, no persona. 171 words against 1,079 in the skill file it is compared with, and a ruleset is reread on every call your agent makes.
Deliver the whole outcome that was asked for, at the level of polish it was asked for. Never shrink the request, ship a reduced version, or stop at a demo. An open request ("complete", "polished", "something I can use every day") is scope: a finished product has every core flow and the features regular users of the best tools of its kind rely on, editing and deleting, undo and recovery, ways to organize and find things, keyboard and screen-reader use, phone layout, dark mode, reduced motion, empty and error states, and keeps the user's data safe, portable and versioned so a later release can still read it.
Spend nothing on anything else. No abstraction with one user, no developer-facing option or config nobody asked for, no wrapper, no speculative extension point, no dependency for a few lines, no duplicated logic, no comment that restates the code, no plan, no narration. Look once, write each file once, verify once.
Final message: what was built and what was not verified, in three sentences.
Install
One file of rules. Nothing runs on your machine.
/plugin marketplace add Obed0101/whittle
/plugin install whittle@whittle
Codex, Pi, Cursor, Windsurf, Cline, Kiro, Copilot, Gemini CLI and anything that reads AGENTS.md: one line each in INSTALL.md. Or paste the three paragraphs into your agent's rules file.
Read before you quote us
Every run, patch and score is in the repository. If your numbers disagree, send the runs. The full list of limits is in the README.
- Ponytail is still cheaper.Whittle costs about a third more per run. What it does not do is pay for the saving with your product.
- Cheaper than no skill on average, not every time.The cost of no skill moved by a third between rounds while nothing in the benchmark changed. On the 14 October kanban runs Whittle cost slightly more.
- Measured on Claude.With GPT-6 Luna, four spec-driven tasks come out about the same for every skill, and two Whittle runs shipped a CLI that crashes on start. Only the open kanban brief separates them.
- Breadth, not quality.On the open briefs a check shows that a feature was built, not how well it works.
- Small sample.Two rounds of five runs per cell, one Claude model, two briefs. On kanban the gap to Ponytail does not overlap; on notes it does in the second round.