pytest-prompts

pytest for LLM prompts. Write tests, detect regressions, run in CI.

from pytest_prompts import prompt_test

@prompt_test()
def test_summary_is_concise(runner):
    result = runner.run(
        prompt="prompts/summarizer.txt",
        input="The Eiffel Tower is a wrought-iron lattice tower in Paris...",
    )
    verdict = runner.judge(result, "The summary is under 50 words and factually correct")
    assert verdict.verdict, verdict.reasoning

$ pytest-prompts run
                                pytest-prompts results
 Test                                           Model             Tokens  Latency  Status
 tests/test_summarizer.py::test_summary...      claude-sonnet-4-6    174    1.2s    PASS
 tests/test_summarizer.py::test_qa_knows...     claude-sonnet-4-6     48    0.9s    PASS

2 passed, 0 failed — 222 tokens total — $0.0009

Change a prompt → pytest-prompts diff main → see exactly what regressed.

Install

uv add pytest-prompts
export ANTHROPIC_API_KEY=sk-ant-...

Quickstart

# 1. Write a test
cat > tests/test_summarizer.py << 'EOF'
from pytest_prompts import prompt_test

@prompt_test()
def test_summary_is_concise(runner):
    result = runner.run(
        prompt="prompts/summarizer.txt",
        input="Long article here...",
    )
    verdict = runner.judge(result, "The summary is under 50 words")
    assert verdict.verdict, verdict.reasoning
EOF

# 2. Run
pytest-prompts run

# 3. Change your prompt, then detect regressions
pytest-prompts diff main

Write a test

Any pytest file. Decorate with @prompt_test(), declare a runner fixture, assert on the result.

from pytest_prompts import prompt_test

@prompt_test(model="claude-sonnet-4-6")
def test_json_extraction(runner):
    result = runner.run(
        prompt="prompts/extractor.txt",
        input='{"name": "Alice", "age": 30}',
    )
    import json
    data = json.loads(result.output)
    assert data["name"] == "Alice"
    assert result.tokens_used < 500

result exposes output, input_tokens, output_tokens, tokens_used, latency_ms, model, cost_usd.

LLM-as-judge

String matching is fragile — use runner.judge() to evaluate outputs semantically:

@prompt_test()
def test_qa_knows_capital_of_france(runner):
    result = runner.run(prompt="prompts/qa.txt", input="What is the capital of France?")
    verdict = runner.judge(result, "The answer correctly identifies Paris as the capital of France")
    assert verdict.verdict, verdict.reasoning

verdict exposes verdict (bool), reasoning (one sentence), criterion, input_tokens, output_tokens, cost_usd. Judge calls are recorded in snapshots alongside the run result.

Detect regressions

# One command — runs tests on main, then on HEAD, compares automatically
pytest-prompts diff main

# Scope to a specific test directory
pytest-prompts diff main --path tests/prompts/

Output:

pytest-prompts diff
 Test                             main           HEAD           Status
 test_summary_is_concise          ✓ 342t 1.2s    ✓ 891t 3.1s    REGRESSION
 test_qa_knows_capital_of_france  ✓ 48t 0.9s     ✓ 48t 0.8s     ok

❌ REGRESSION  test_summary_is_concise
   tokens 342 → 891 (+160%)

Exit code 1 on any regression. Wire it into CI and you're done.

CI (GitHub Actions)

- uses: actions/checkout@v4
  with:
    fetch-depth: 0  # required for base-ref diff mode
- uses: chahine-tech/pytest-prompts@v0.1
  with:
    path: tests/prompts
    anthropic-api-key: ${{ secrets.ANTHROPIC_API_KEY }}
    base-ref: main
    github-token: ${{ secrets.GITHUB_TOKEN }}

On pull requests, the action runs your tests against both main and your branch, compares them, fails the job on regressions, and posts a summary comment on the PR.

Outputs: passed, failed, total-tokens, total-cost-usd, regressions.

What's included

• @prompt_test decorator with pytest integration
• Runner for the Anthropic API (Claude Sonnet 4.6 default)
• runner.judge() — LLM-as-judge for semantic assertions
• pytest-prompts run — run tests, summarize tokens/latency/cost
• pytest-prompts diff <ref> — run on a git ref + HEAD, detect regressions automatically

Not here yet: OpenAI/Gemini adapters, static prompt analysis, HTML reports. Open an issue — priorities come from usage, not a roadmap.

License

MIT.