Sharpen Claude Prompt Refiner

sharpen

A Claude skill that rewrites a vague prompt into a tight spec before any work starts, then makes the model take the shortest correct path to it.

Built to fix four things that waste the most tokens:

1. Wrong target — the work is correct but not what you wanted
2. Invented facts — confident output with nothing checkable under it
3. Wasteful path — the right answer reached the long way round
4. Default AI look — correct, efficient, and instantly recognisable as machine-made

What it does

Before touching the task, Claude reduces your request to five lines and shows them, then starts work in the same reply:

If something is missing, it looks it up rather than asking. It only asks when a wrong guess would waste the whole run, and never more than two questions.

For anything visual it commits to a stated design direction and runs a slop audit before shipping: no purple gradients, no glassmorphism, no emoji bullets, no grids of identical rounded cards, no "unlock" or "seamless".

Install

**Claude Code**

git clone https://github.com//sharpen.git ~/Projects/sharpen
mkdir -p ~/.claude/skills
ln -sfn ~/Projects/sharpen ~/.claude/skills/sharpen

**Claude app** — open Skills settings and add `SKILL.md`.

Use

It triggers on its own for vague, multi-part or visual requests. To force it:

Also fires on "refine my prompt", "be efficient", "don't hallucinate", "no AI slop".

Not for

Small, unambiguous questions. A one-line answer should stay a one-line answer, and adding a spec block to trivial work is the same waste the skill exists to prevent.

License

MIT