Superb Prompt — Code Quality skill for Claude Code

superb-prompt

A Claude skill that builds or fixes prompts using the five-part framework (role, task, context, constraints, format), names the reader, asks for one insight, and checks every prompt against a nine-point "superb test" before it goes out.

A good prompt gets a correct answer. A superb prompt also gets an answer that is written for a named reader and tells that reader something they didn't already know. Every prompt this skill writes does both.

What it does

• Fills in the five parts from what you said, what you attached, and what Claude knows of your work: role, task, context, constraints, format. Anything it can't know gets a [BRACKETED] placeholder rather than an invented fact.
• Names the audience in a way that changes how the AI writes: their title, what they already know, and the decision they face. "For leadership" is weak; "for a health plan Chief Dental Officer deciding which coordination program to fund next year" is strong.
• Adds exactly one insight request, the "so what" that pushes past a summary into judgment. Reformat prompts get none, on purpose: a formatting prompt should change only the layout.
• Matches the prompt type, each with its must-haves and a template: summarize, role-based, reformat, compare/contrast, evaluate. Real tasks often combine two.
• Runs the superb test before delivering: task verb, role, audience, content, format, constraints, insight, accuracy anchoring (citations required whenever facts, numbers or policy are involved), length.
• Pre-empts weak output with two or three follow-ups drawn from a diagnosis table: too generic, off-tone, hard to use, hallucinated, shallow, wrong reader, and the fix for each.

Four modes

Mode	What you give it	What comes back
A. Build	A rough request	A quick (1–2 sentence) prompt, or a full version in two forms: structured (five labeled parts) and narrative (2–3 paragraphs)
B. Review	An existing prompt, optionally with