Using Prompt Lab

Using Prompt-Lab

Overview

Prompt-lab is a CLI tool for testing prompt variants across LLM providers using LLM-as-judge evaluation.

system.md (optional) + prompt.md + inputs.yaml → LLM → response → judge.md → score

Create multiple variants (v1, v2, ...) to A/B test different prompt approaches against the same inputs and judge, then compare with statistical significance.

Experiment Structure

experiments/
  my-experiment/
    experiment.md       # Config: name, models, runs (required)
    judge.md            # Scoring rubric (required)
    inputs.yaml         # Shared test cases (optional)
    v1/                 # Variant (at least one required)
      prompt.md         # User message (required)
      system.md         # System message (optional)
      tools.yaml        # Tool definitions (optional)
    v2/                 # Another variant to compare
      prompt.md

Fallback resolution: judge.md and inputs.yaml are checked in the variant directory first, then the experiment directory. This enables shared test cases while allowing per-variant overrides.

File Quick Reference

For detailed file formats and judge rubric design, see experiment-reference.md.

Creating Experiments

From config file (recommended)

prompt-lab new --config spec.yaml

Spec format:

name: my-experiment
description: What this tests
hypothesis: Expected outcome
models:
  - openai:gpt-4o-mini
  - anthropic:claude-sonnet-4-20250514
runs: 5
path: experiments
key_refs:
  openai: MY_CUSTOM_OPENAI_KEY

inputs:
  - id: case-1
    field_name: value1
  - id: case-2
    field_name: value2

judge:
  model: openai:gpt-4o
  score_range: [0, 5]
  temperature: 0
  chain_of_thought: true
  rubric: |
    Your rubric with {{ prompt }} and {{ response }}.

variants:
  v1:
    prompt: |
      Prompt template with {{ field_name }}.
    system: |
      Optional system prompt.
  v2:
    prompt: |
      Alternative approach with {{ field_name }}.

Interactive wizard

prompt-lab new

Manual creation

Create the directory structure and files by hand. Best for complex experiments.

Running Experiments

# Run all variants in an experiment
prompt-lab run experiments/my-experiment

# Run a single variant
prompt-lab run experiments/my-experiment/v1

# Run specific model only
prompt-lab run experiments/my-experiment/v1 --model openai:gpt-4o-mini

# Skip cache (fresh API calls)
prompt-lab run experiments/my-experiment --no-cache

# Hide progress bar
prompt-lab run experiments/my-experiment -q

# Custom API key env var (format: provider:ENV_VAR)
prompt-lab run experiments/my-experiment -k openai:MY_OPENAI_KEY

What happens during a run

For each (input, run_number, model) combination, concurrently:

1. Provider renders prompt.md + system.md with input variables via Jinja2
2. LLM generates a response (with optional tool calls)
3. Judge LLM scores the response using judge.md rubric
4. Result saved as JSON under variant/results/{timestamp}/responses/

Cache is automatically disabled when runs > 1 to ensure independent responses.

Viewing Results

Results table

prompt-lab results experiments/my-experiment/v1

Shows per-input scores with mean, 95% confidence interval, and score range.

# View a specific historical run
prompt-lab results experiments/my-experiment/v1 --run 2026-01-25T19-30-00

Detailed responses with judge reasoning

# All responses
prompt-lab show experiments/my-experiment/v1

# Filter by input
prompt-lab show experiments/my-experiment/v1 --input alice

# Filter by model
prompt-lab show experiments/my-experiment/v1 --model openai:gpt-4o-mini

# Combine filters
prompt-lab show experiments/my-experiment/v1 --input alice --model openai:gpt-4o-mini

# Specific historical run
prompt-lab show experiments/my-experiment/v1 --run 2026-01-25T19-30-00

Results storage

variant/results/{timestamp}/
  run.yaml                              # Run metadata (duration, models, counts)
  stats.yaml                            # Per-input stats (mean, CI, stddev, scores)
  responses/
    {input_id}_run{N}_{provider}-{model}.json   # Individual result

Each response JSON contains: input_id, model, run_number, cached, latency_ms, input_tokens, output_tokens, response (content + tool_calls), judge (score + reasoning).

Comparing Variants

prompt-lab compare experiments/my-experiment

Shows comparison table across all variants:

• Mean score per variant with 95% confidence intervals
• Average latency
• Total runs
• Statistical significance via Welch's t-test (p-value)
• Experiment hypothesis

Tells you whether v1 is actually better than v2, or if the difference is just noise.

Cleaning Up

# Clean single variant results
prompt-lab clean experiments/my-experiment/v1

# Clean all variants in an experiment
prompt-lab clean experiments/my-experiment

# Skip confirmation
prompt-lab clean experiments/my-experiment --yes

Cache Management

prompt-lab cache clear

Cache stores LLM responses to avoid redundant API calls during development. Automatically disabled when runs > 1.

Experiment Design Tips

• Variant strategy: Each variant tests a different prompt approach (zero-shot vs few-shot, formal vs casual, structured vs freeform, with/without CoT)
• Statistical reliability: Use runs: 5+ for meaningful confidence intervals
• Judge selection: Use a different model family as judge to reduce self-enhancement bias, or multi-judge with models: (plural) in judge.md
• Model format: Always provider:model (e.g., openai:gpt-4o-mini, anthropic:claude-sonnet-4-20250514)
• Supported providers: openai:*, anthropic:*

Common Mistakes