Prompt Refiner

What this skill does

This skill takes a rough or underperforming AI prompt and rewrites it to be clearer, more specific, and more likely to produce consistent, high-quality output from a language model. It identifies the root causes of vague or ineffective prompts — missing context, absent output format specs, ambiguous instructions — and systematically addresses each one. It also explains every change so you can learn the principles, not just get a one-time fix.

Use this when your prompts produce inconsistent results, when the model frequently misunderstands what you want, or when you're building a prompt that will run in production.

How to use

Claude Code / Cline

Copy this file to .agents/skills/prompt-refiner/SKILL.md in your project root.

Then paste your prompt and ask:

• "Use the Prompt Refiner skill to improve this prompt."
• "Refine this system prompt using the Prompt Refiner skill — it's producing inconsistent outputs."

Include what the prompt is for (the model it targets, the task, what's going wrong with current outputs) to get more targeted improvements.

Cursor

Add the "Prompt / Instructions" section to your .cursorrules file. Paste your prompt into the chat and ask for a refinement.

Codex

Paste your original prompt and a description of what output you're trying to achieve, then include the instructions below.

The Prompt / Instructions for the Agent

When asked to refine a prompt, follow these steps:

1. Analyze the original prompt for these weaknesses:

• Vague task definition — the model has to guess what "good" looks like
• No role assignment — not telling the model what persona or expertise to adopt
• Missing output format — no specification of length, structure, format (JSON, markdown, bullet list, etc.)
• Missing context — no background about the user, use case, or constraints
• Ambiguous pronouns or references — "it", "this", "the thing" without clear antecedents
• Conflicting instructions — asking for brevity and comprehensiveness in the same breath
• No examples — complex tasks benefit from at least one example of desired input/output
• No constraints — no guidance on what to avoid or exclude
• Negative-only instructions — "don't use jargon" without specifying the preferred alternative

1. Identify which weaknesses are present in the original prompt. Note each one specifically.

1. Rewrite the prompt applying these prompt engineering best practices:

• Assign a clear role: Start with "You are a [specific expert/persona]..." when helpful
• Define the task precisely: Use action verbs and be specific about the deliverable
• Specify the output format: Length, structure, tone, and file format if relevant
• Provide relevant context: Who is asking, what platform will this run on, what the output will be used for
• Add constraints: What to avoid, what assumptions to make, what topics are out of scope
• Include an example (few-shot) if the task is complex or the output format is unusual
• Use numbered steps for multi-part tasks rather than a wall of instructions
• Separate system instructions from user content using clear delimiters when relevant

1. Explain every change in a "Changes Made" section. For each change:

• Name the principle applied
• Explain what was missing in the original
• Explain how the change addresses it

1. Optionally provide variations if different use cases call for different versions (e.g., a terse version vs. a detailed version, or a version for GPT-4o vs. Claude 3).

1. Output format:

## Original Prompt Analysis
**Weaknesses identified:**
- [Weakness 1]: [Explanation]
- [Weakness 2]: [Explanation]
...

---

## Refined Prompt

[The full rewritten prompt]

---

## Changes Made

### 1. [Change title]
**Principle:** [Prompt engineering principle applied]
**Before:** "[relevant original text or "(missing)"]"
**After:** "[relevant new text]"
**Why:** [Explanation]

### 2. ...

---

## Optional Variations
[If applicable, a shorter version or model-specific variant]

Example

Input to Agent:

"Use the Prompt Refiner skill on this prompt: 'Summarize this article. Make it good and clear. Don't be too long.'"

Output from Agent:

## Original Prompt Analysis Weaknesses identified: - No role assignment: The model has no persona to adopt, so it defaults to a generic voice. - Vague quality standard: "Make it good and clear" gives the model nothing concrete to optimize for. - Ambiguous length constraint: "Don't be too long" is subjective — 3 sentences? 3 paragraphs? - Missing output format: No specification of structure (paragraph, bullet points, TL;DR format, etc.). - Missing audience context: Who will read this summary? A technical reader? A general audience? - No constraints on what to include: Should the summary include the author's conclusion? Statistics? Quotes?

>

---

>

## Refined Prompt

>

``` You are a professional editor writing for a general audience with no prior knowledge of the topic.

>

Summarize the article below in exactly 3–5 sentences. Your summary must: - Cover the main argument or finding - Include the single most important supporting fact or statistic - End with the article's conclusion or key takeaway - Use plain language — avoid jargon and technical terms - Not exceed 100 words

>

Article: """ [INSERT ARTICLE TEXT HERE] """ ```

>

---

>

## Changes Made

>

### 1. Added role assignment Principle: Role prompting anchors tone, vocabulary, and expertise level. Before: (missing) After: "You are a professional editor writing for a general audience..." Why: The model now has a consistent voice and knows the target reader, which shapes word choice and complexity.

>

### 2. Specified exact length Principle: Concrete constraints produce consistent outputs. Before: "Don't be too long" After: "exactly 3–5 sentences" and "Not exceed 100 words" Why: Two complementary length caps prevent both a single bloated sentence and five dense run-ons.

>

### 3. Defined required content Principle: Explicit content requirements reduce model guessing. Before: "Make it good and clear" After: Numbered list of what must appear in the summary Why: The model no longer decides what "good" means — the criteria are explicit.

>

### 4. Added a text delimiter Principle: Delimiters prevent prompt injection and clarify input boundaries. Before: (implied the article follows) After: """ [INSERT ARTICLE TEXT HERE] """ Why: The triple-quote delimiter makes it unambiguous where the instructions end and the article begins.