prompt-prof

Analyze your AI coding assistant prompts to become a better prompter.

Track prompt quality, identify patterns, and get actionable insights to improve your effectiveness with Claude Code and Cursor.

Why prompt-prof?

AI coding assistants are powerful, but their effectiveness depends heavily on how you prompt them. This tool helps you:

• Understand your prompting patterns - See what types of prompts you use most
• Identify inefficiencies - Find vague prompts, excessive retries, and clarifications
• Track quality over time - Monitor your average prompt scores
• Learn from your best prompts - See what works and replicate it

Supported AI Assistants

| Assistant | Data Location | What's Analyzed | |-----------|---------------|-----------------| | Claude Code | ~/.claude/ | Sessions, costs, tokens, tool usage | | Cursor | ~/.cursor/ | Agent transcripts, code generations |

Installation

npm install -g prompt-prof

Quick Start

# Comprehensive report for all AI assistants
prompt-prof report

# Today only
prompt-prof report -d 1

# Past 30 days
prompt-prof report -d 30

Example Output

╔════════════════════════════════════════════════════╗
║    Prompt Effectiveness Report (Past 7 days)     ║
╚════════════════════════════════════════════════════╝

🤖 CLAUDE CODE

┌───────────────┬────────┐
│ Total Prompts │ 76     │
│ Average Score │ 59/100 │
│ Sessions      │ 10     │
└───────────────┴────────┘

Prompt Types:
┌──────────────────┬───────┬───────┐
│ Type             │ Count │ %     │
├──────────────────┼───────┼───────┤
│ Code Generation  │ 4     │ 5.3%  │
│ Questions        │ 21    │ 27.6% │
│ Commands/Actions │ 16    │ 21.1% │
│ Clarifications   │ 5     │ 6.6%  │
│ Other            │ 27    │ 35.5% │
└──────────────────┴───────┴───────┘

Score Distribution:
┌─────────────────┬───────┬───────┐
│ Rating          │ Count │ %     │
├─────────────────┼───────┼───────┤
│ Excellent (90+) │ 0     │ 0.0%  │
│ Good (70-89)    │ 1     │ 1.3%  │
│ Fair (50-69)    │ 75    │ 98.7% │
│ Poor (<50)      │ 0     │ 0.0%  │
└─────────────────┴───────┴───────┘

📝 CURSOR

┌───────────────┬────────┐
│ Total Prompts │ 43     │
│ Average Score │ 61/100 │
│ Transcripts   │ 3      │
└───────────────┴────────┘
...

What the Report Shows

Prompt Types

| Type | Examples | |------|----------| | Code Generation | "Create a function...", "Implement...", "Build a component..." | | Questions | "What is...", "How do I...", "Why does...", "Can you explain..." | | File Operations | "Read file...", "Find files matching...", "Search for..." | | Commands/Actions | "Run tests", "Fix the bug", "Update the config", "Deploy" | | Clarifications | "No, I meant...", "Actually...", "Not that, the other one" | | Other | Everything else |

Score Ratings

| Rating | Score | What it Means | |--------|-------|---------------| | Excellent | 90-100 | Specific, clear context, actionable | | Good | 70-89 | Solid prompts with minor improvements possible | | Fair | 50-69 | Works but could be more specific | | Poor | 0-49 | Vague, missing context, or retries |

All Commands

Global (Both Claude Code & Cursor)

prompt-prof report              # Comprehensive report (default: 7 days)
prompt-prof report -d 1         # Today only
prompt-prof report -d 30        # Past 30 days
prompt-prof stats               # Quick stats overview

Claude Code

# Reports
prompt-prof claude report daily              # Today's summary
prompt-prof claude report daily -d 2026-01-22  # Specific date
prompt-prof claude report weekly             # Past 7 days

# Find Best & Worst Prompts
prompt-prof claude patterns best             # Top 10 prompts
prompt-prof claude patterns best -n 20       # Top 20 prompts
prompt-prof claude patterns worst            # Bottom 10 prompts
prompt-prof claude patterns worst -d 30      # Look back 30 days

# Cost Analysis
prompt-prof claude cost summary              # Cost breakdown
prompt-prof claude cost summary -d 30        # Past 30 days

# Sessions
prompt-prof claude list sessions             # Recent sessions
prompt-prof claude list sessions -n 20       # Show 20 sessions
prompt-prof claude analyze <sessionId>       # Analyze specific session

Cursor

# Reports
prompt-prof cursor report summary            # Usage summary (default: 7 days)
prompt-prof cursor report summary -d 30      # Past 30 days

# Find Best & Worst Prompts
prompt-prof cursor patterns best             # Top 10 prompts
prompt-prof cursor patterns worst            # Bottom 10 prompts

# Transcripts
prompt-prof cursor list transcripts          # Recent transcripts
prompt-prof cursor analyze <transcriptId>    # Analyze specific transcript

# Database Stats
prompt-prof cursor stats                     # Code generations, models, file types

Quality Scoring System

Each prompt is scored 0-100 based on 4 dimensions:

Clarity (25%)

| Factor | Points | |--------|--------| | File/function references | +10 | | Clear action verbs (create, fix, update) | +8 | | Line numbers, error messages, code snippets | +5 | | Vague commands ("fix it", "make it work") | -15 | | Short prompts without context | -10 |

Context (25%)

| Factor | Points | |--------|--------| | Proper follow-up references | +10 | | References recent tool output | +8 | | Cold start with unclear reference | -15 | | Assumes missing context | -10 |

Efficiency (25%)

| Factor | Points | |--------|--------| | Retry of previous prompt (>60% similar) | -20 | | Excessive verbosity (500+ words) | -10 |

Outcome (25%)

| Factor | Points | |--------|--------| | High tool success rate | +10 | | Required clarification follow-up | -15 | | Conversation loop detected | -20 |

Data Sources

The tool reads local data from your AI assistants:

Claude Code (~/.claude/)

• projects/[path]/[sessionId].jsonl - Full conversation history
• stats-cache.json - Aggregated usage statistics

Cursor (~/.cursor/)

• projects/[project]/agent-transcripts/*.txt - Agent conversations
• ai-tracking/ai-code-tracking.db - SQLite database with code generation stats

Note: All analysis happens locally. No data is sent anywhere.

Programmatic Usage (SDK)

import {
  // Claude Code
  parseSessions,
  analyzeSessionPrompts,
  calculateAverageQuality,

  // Cursor
  parseCursorTranscripts,
  analyzeCursorTranscriptPrompts,

  // Classification
  classifyPrompt,
  calculatePromptStats,
} from 'prompt-prof';

// Analyze Claude Code sessions
const sessions = await parseSessions(startDate, endDate);
for (const session of sessions) {
  const analyzed = analyzeSessionPrompts(session);
  console.log(`Avg quality: ${calculateAverageQuality(analyzed)}`);
}

// Analyze Cursor transcripts
const transcripts = parseCursorTranscripts(startDate, endDate);
for (const transcript of transcripts) {
  const analyzed = analyzeCursorTranscriptPrompts(transcript);
  // Process results...
}

// Classify a single prompt
const type = classifyPrompt("Create a REST API endpoint for users");
// Returns: 'code_generation'

Tips for Better Prompts

Based on the scoring system:

1. Be specific - Include file names, function names, line numbers

   • Bad: "Fix the bug"
   • Good: "Fix the TypeError in src/api/users.ts:42"

2. Use action verbs - "Create", "Fix", "Update", "Refactor"

   • Bad: "The login doesn't work"
   • Good: "Fix the login validation that rejects valid emails"

3. Provide context - Include error messages, expected behavior

   • Bad: "It's broken"
   • Good: "Getting 'Cannot read property of undefined' when clicking submit"

4. Avoid vague language - Skip "fix it", "make it work", "do it"

   • Bad: "Do the same thing for the other file"
   • Good: "Apply the same validation logic to src/api/orders.ts"

5. Reduce clarifications - Be specific upfront

   • If you find yourself saying "No, I meant..." often, add more detail initially

Contributing

Issues and PRs welcome at github.com/sanketagarwal/prompt-prof

License

MIT