Skip to content

Recipe: turn a bad prompt into a great one — automatic prompt scoring + repair with Claude #798

Description

@Amritha902

The gap

Every new Claude user writes weak prompts at first — but the hard part isn't
fixing them, it's knowing what's wrong. The cookbook has great evaluation
recipes, but none that take a messy real-world prompt, diagnose it, and repair
it. I'd love to contribute that recipe.

What it does (before → after, in one notebook)

Input: "fix my resume idk make it good"
Output: a scored diagnosis + a rewritten, structured prompt that measurably
scores higher.

Concrete outline

  1. Score — Claude grades a draft against 5 best-practice dimensions
    (role, context, specificity, format, structure) and returns a structured
    JSON rubric with a per-dimension reason for any low score.
  2. Repair — a single "prompt compiler" call rewrites the draft into a
    well-structured prompt, preserving the user's intent.
  3. Prove it worked — re-score the rewritten prompt and show the lift; run
    it over a small batch to report average improvement (a real eval loop, not
    a vibe check).
  4. Make it reusable — wrap it in a drop-in score() / improve() helper
    readers can paste into their own apps.

Why it's worth merging

  • Fills a genuine gap (diagnose-and-repair, not just evaluate).
  • Immediately practical — anyone building on Claude wants "make my users'
    prompts better."
  • Self-contained, matches your existing notebook style, easy to review.

Why I can deliver it cleanly

I already built these exact ideas into a working, open-source browser
extension — with 100+ passing tests, reproducible benchmarks, and design docs:
https://github.com/Amritha902/promptcomplete
I'll adapt the concepts into a polished, standalone notebook.

Try it live (no install): https://claude.ai/code/artifact/4d15fab9-3403-4cba-b916-a4442b0768b7

Does this fit, and is anything similar already planned? Happy to tune the
scope to whatever you'd find most useful. I can have a draft ready quickly.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions