promptlab ⚡

Automated testing for LLM prompts. Write test cases in YAML, run them against Claude or OpenAI, get pass/fail results in your terminal.

Like pytest, but for prompts.

pip install promptlab-cli
promptlab run tests/

✅ summarize_article :: returns_short_summary        PASS (1.2s)
✅ summarize_article :: mentions_key_points           PASS (1.1s)
❌ translate_text :: preserves_tone                   FAIL (0.9s)
   Expected: contains "formal"
   Got: "Here is the translated text in a casual style..."

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Results: 2 passed, 1 failed, 3 total (3.2s)

Why?

You're building an app with Claude or GPT. Your prompt works today. Tomorrow you tweak it and something breaks. You don't notice until a user complains.

promptlab catches prompt regressions before they ship. Define what good output looks like, run tests on every change, and know immediately if something broke.

Quickstart

Install

pip install promptlab-cli

Set your API key

export ANTHROPIC_API_KEY=sk-ant-...
# or
export OPENAI_API_KEY=sk-...

Write a test file

Create tests/summarize.yaml:

prompt: |
  Summarize this article in 2-3 sentences:
  {{ article }}

model: claude-sonnet-4-20250514

tests:
  - name: short_summary
    vars:
      article: |
        The Federal Reserve held interest rates steady on Wednesday,
        keeping the benchmark rate in the 5.25%-5.50% range. Chair
        Jerome Powell said the committee needs more confidence that
        inflation is moving toward the 2% target before cutting rates.
    assert:
      - type: max_tokens
        value: 100
      - type: contains
        value: "Federal Reserve"
      - type: contains
        value: "interest rate"

  - name: handles_empty_input
    vars:
      article: ""
    assert:
      - type: not_contains
        value: "error"
      - type: min_length
        value: 10

Run it

promptlab run tests/

Test File Format

Each .yaml file defines a prompt and its test cases:

# The prompt template. Use {{ variable }} for inputs.
prompt: |
  You are a helpful assistant. {{ instruction }}

# Which model to use
model: claude-sonnet-4-20250514   # or gpt-4o, claude-haiku-4-5-20251001, etc.

# Optional system prompt
system: "You are a concise technical writer."

# Optional model parameters
temperature: 0
max_tokens: 500

# Test cases
tests:
  - name: test_name
    vars:
      instruction: "Explain what a CPU does in one sentence."
    assert:
      - type: contains
        value: "processor"
      - type: max_tokens
        value: 50

Assertion Types

LLM-as-Judge

The most powerful assertion type. Uses a second LLM call to evaluate output quality:

tests:
  - name: helpful_response
    vars:
      question: "How do I fix a memory leak in Python?"
    assert:
      - type: llm_judge
        value: "Does this response provide specific, actionable debugging steps? Answer YES or NO."

Supported Models

Anthropic (Claude):

• claude-sonnet-4-20250514
• claude-haiku-4-5-20251001
• Set ANTHROPIC_API_KEY environment variable

OpenAI:

• gpt-4o
• gpt-4o-mini
• Set OPENAI_API_KEY environment variable

CLI Commands

# Run all test files in a directory
promptlab run tests/

# Run a single test file
promptlab run tests/summarize.yaml

# Verbose output (show full LLM responses)
promptlab run tests/ --verbose

# Output results as JSON
promptlab run tests/ --json

# Dry run (show what would be tested without calling APIs)
promptlab run tests/ --dry-run

Use Cases

• Prompt regression testing — Run tests in CI/CD to catch regressions
• Prompt comparison — Test the same cases across different models
• Guard rails validation — Verify your prompt rejects harmful inputs
• Output format checking — Ensure structured output matches expectations

Development

git clone https://github.com/vigp17/promptlab.git
cd promptlab
pip install -e ".[dev]"
pytest

Roadmap

• YAML test definitions
• Claude and OpenAI support
• 9 assertion types including LLM-as-judge
• CLI with colored output
• Cost tracking per test run
• HTML report generation
• Parallel test execution
• GitHub Actions integration
• Prompt versioning and diff
• Custom scoring functions

License

MIT