sharpenlisted
Install: claude install-skill sazzadurrahmaan/sharpen
# Sharpen
Turn a rough request into a prompt an agent can execute correctly on the first try.
**Golden rule:** do not tell the agent what you assume the solution is. Tell it what
outcome is required, what evidence to inspect, what must not break, and let the
existing codebase determine the implementation.
Priority when these conflict: **correctness > intent preservation > execution
quality > token efficiency**. Never trade correctness for brevity.
## Procedure
### 1. Classify
Pick one: bug fix / debugging / feature / refactor / code review / review-feedback
resolution / research / UI / performance / security / testing / docs / infra / mixed.
The class decides which instructions are worth including. Skip the rest.
Class-specific guidance lives in `references/by-task-type.md` — read that file only
for the class you picked, and only when the task is non-trivial.
### 2. Extract the signal
From the user's text, pull out:
| Slot | Question |
| --- | --- |
| Objective | What outcome do they actually want? |
| Current state | What exists now? |
| Problem | What is wrong, missing, or unclear? |
| Evidence | Errors, logs, paths, screenshots, repro steps they gave |
| Constraints | What must not change or break |
| References | Files, branches, PRs, issues, URLs, designs |
Drop greetings, filler, repetition, frustration, self-explanation, and generic
engineering advice. Keep every concrete token: file paths, symbol names, routes,
error strings, branch names, PR numbers, versi