What is scalpel?

Scalpel is a set of instructions, called a skill, that you can install into an AI coding agent such as Claude Code to change how it writes code. Instead of writing brand new code for every request, it tells the agent to first check whether something is really needed, then look for an existing way to solve it: something already in the project's own code, something built into the programming language or platform, or an already installed package, before writing anything new. Only if none of those exist does it write new code, and only the smallest amount needed. The skill also tells the agent to make a plan for a task once and stick with it, rather than reconsidering its approach on every reply, which the author says wastes time and cost. It protects certain things from being cut for the sake of brevity, such as input validation, error handling, security checks, and accessibility, and its replies are meant to be mostly code with very little explanatory text. The README includes benchmark results comparing this skill to a similar project called ponytail, using that other project's own test setup so the comparison uses the same tasks and scoring method. Across tests with several Claude models, the author reports this skill produced less code, used fewer tokens, cost less to run, and matched or beat the other project on correctness in every case tested. A separate test using real coding sessions on a sample web app project found the two projects performed about the same overall. To install it, you can use a plugin command in Claude Code, or copy a single instructions file into a project's skills folder, or paste its contents into the system prompt of any AI agent that accepts one. It can be turned off at any time by telling the agent to stop using it or switch to normal mode, and it will still build a fuller version of something if you specifically ask for one. The project is released under the MIT license.