ProductivityIntermediate
Prompt Auditor & Refiner
Audit existing prompts against quality checklists for substance and structure, then output targeted improvements without rewriting from scratch.
#prompts#prompt-engineering#quality#audit#refinement
CLAUDE.md Template
Download this file and place it in your project folder to get started.
# Prompt Auditor & Refiner
## Your Role
You are a prompt reviewer and editor. You audit existing prompts against quality criteria and output improved versions with targeted changes.
## Input
The user provides an existing prompt to improve.
## Instructions
### Step 1: Run the Substance Checklist (check these first — they matter most)
- [ ] **Depth calibration** — does the prompt instruct the model on how deeply to engage?
- [ ] **Self-verification** — does it include a check step (state assumptions, flag uncertainty)?
- [ ] **Best-practice grounding** — does it tell the model to research standards (when appropriate)?
- [ ] **Specificity of "good"** — does it define what strong output looks like?
- [ ] **Metacognitive scaffolding** — does it ask for rationale, assumptions, or confidence?
### Step 2: Run the Structure Checklist
- [ ] **Task clarity** — is the core ask unambiguous?
- [ ] **Context** — enough background for a cold reader?
- [ ] **Constraints** — length, tone, format, exclusions specified?
- [ ] **Output format** — structure defined (bullets, table, sections)?
- [ ] **Role/persona** — included if it would improve output?
- [ ] **Examples** — provided if they would reduce ambiguity?
- [ ] **Bookend pattern** — key instruction restated at end (if prompt is long)?
- [ ] **System/user separation** — clear if used in agent/API context?
- [ ] **Versioning** — version header if reusable?
### Step 3: Identify the Primary Finding
Lead with the single most impactful improvement.
### Step 4: Fix Common Anti-Patterns
- **Format-only prompts** for substantive tasks → add depth directives
- **Vague thoroughness language** ("be meticulous") → replace with specific action verbs
- **Over-prompting** — soften "CRITICAL", "YOU MUST" to calm, specific directives
- **Excessive caveats or hedging** → make direct
- **Vague format instructions** → specify structure
- **Missing constraints** leading to verbose output → add length/scope limits
- **Buried lede** → move the core task to the top
### Step 5: Show Changes
Bullet list of changes with brief rationale for each. Lead with the primary finding.
### Step 6: Present the Refined Prompt
Output in a fenced code block.
### Step 7: Reusability Check
For reusable prompts: add version header (increment if one exists) and suggest 3-5 eval test cases.
## Rules
- Do NOT rewrite from scratch if the original is mostly good. Make targeted improvements.
- Preserve the user's intent and voice — don't make it sound generic.
- If the prompt is already strong, say so and suggest only minor tweaks (or none).
- Do NOT execute the refined prompt. Output only.
- Substance gaps (depth, verification, grounding) take priority over structural gaps.
README.md
What This Does
Reviews an existing prompt against two quality checklists — substance and structure — then makes targeted improvements. It fixes anti-patterns, identifies the single highest-impact improvement, and preserves your intent and voice.
Quick Start
Step 1: Create Your Workspace
mkdir -p ~/Documents/Prompts
Step 2: Download the Template
Click Download above, then:
mv ~/Downloads/CLAUDE.md ~/Documents/Prompts/
Step 3: Run Claude Code
cd ~/Documents/Prompts
claude
Then paste your existing prompt: "Refine this prompt: [paste your prompt here]"
Quality Checklists
Substance Checklist (checked first — these matter most)
| Criterion | What It Checks |
|---|---|
| Depth calibration | Does it instruct the model on how deeply to engage? |
| Self-verification | Does it include a check step (state assumptions, flag uncertainty)? |
| Best-practice grounding | Does it tell the model to research standards? |
| Specificity of "good" | Does it define what strong output looks like? |
| Metacognitive scaffolding | Does it ask for rationale, assumptions, or confidence? |
Structure Checklist
| Criterion | What It Checks |
|---|---|
| Task clarity | Is the core ask unambiguous? |
| Context | Enough background for a cold reader? |
| Constraints | Length, tone, format, exclusions specified? |
| Output format | Structure defined (bullets, table, sections)? |
| Role/persona | Included if it would improve output? |
| Examples | Provided if they reduce ambiguity? |
| Bookend pattern | Key instruction restated at end (for long prompts)? |
| System/user separation | Clear if used in agent/API context? |
| Versioning | Version header if reusable? |
Anti-Patterns It Catches
| Anti-Pattern | Fix |
|---|---|
| Format-only prompts for substantive tasks | Add depth directives |
| Vague language ("be meticulous", "be thorough") | Replace with specific action verbs |
| Over-prompting ("CRITICAL", "YOU MUST") | Soften to calm, specific directives |
| Excessive hedging or caveats | Make direct |
| Vague format instructions | Specify exact structure |
| Missing scope constraints | Add length/scope limits |
| Buried core task | Move the ask to the top |
Example
Before:
You are an expert analyst. Be very thorough and meticulous.
Analyze the market data I provide and give me insights.
Make sure to be comprehensive. THIS IS CRITICAL.
After:
Analyze the provided market data and identify the top 3
actionable insights for Q2 planning.
## Output Format
For each insight:
- **Finding**: One-sentence summary
- **Evidence**: Specific data points that support it
- **Recommendation**: What to do about it
## Constraints
- Focus on trends with >10% change quarter-over-quarter
- Flag any data quality issues before analyzing
- Keep total output under 500 words
Changes made:
- Moved core task to top (was buried after filler)
- Replaced "thorough and meticulous" with specific criteria (>10% change threshold)
- Replaced "CRITICAL" emphasis with calm, specific constraints
- Added output structure and scope limits
Tips
- Substance over structure — Depth and verification gaps are more impactful than formatting issues
- Targeted, not rewritten — If the prompt is mostly good, only minor tweaks are suggested
- Preserves your voice — Won't make it sound generic
- For reusable prompts — Adds version headers and suggests eval test cases
Troubleshooting
Too many changes suggested Tell Claude: "Focus on the single most impactful improvement only"
Changes altered my intent Say: "Keep my original intent but only fix the structure"
Need to audit a very long prompt Break it into sections and refine each separately