| name | create-eval |
| description | Create a new JudgeJudy evaluation config file interactively |
| user-invocable | true |
| disable-model-invocation | true |
| allowed-tools | Read, Write, Glob, Bash, AskUserQuestion |
| argument-hint | ["output-filename.yaml"] |
Create a JudgeJudy Evaluation Config
You are helping the user create a new evaluation YAML config for JudgeJudy.
Step 1: Gather Basic Info
Ask the user for the following (one question at a time, use AskUserQuestion):
- What are you evaluating? (e.g., "text generation quality", "TTS naturalness", "image generation from prompts", "video generation realism")
- What modality? โ text, image, audio, or video
- Which provider and model for generation? Show them the available options:
- Text: openai (gpt-4o, gpt-4.1), anthropic (claude-sonnet-4-6), google (gemini-2.5-pro/flash), together, ollama
- Image: openai (dall-e-3), wavespeed (seedream-v3.1, seedream-v4)
- Audio: openai (tts-1, tts-1-hd), elevenlabs (eleven_multilingual_v2, eleven_v3), cartesia (sonic-2, sonic-3)
- Video: wavespeed (seedance-v1.5-pro/text-to-video, wan-2.5-t2v), falai (kling-video/v3)
- What should the judge evaluate? Ask them to describe the quality criteria. Examples: "accuracy and clarity", "visual quality and prompt adherence", "naturalness and intelligibility", "realism and temporal consistency"
- Which provider/model for the AI judge? Recommend:
- Text/Image judging: anthropic (claude-sonnet-4-6) or openai (gpt-4o)
- Audio judging: google (gemini-2.0-flash) โ supports native audio
- Video judging: anthropic (claude-sonnet-4-6) โ via frame extraction
- Do you have test cases already, or should I generate some? If they want generated ones, ask how many (default 10) and what topics/themes.
- Do you want automated metrics too? Suggest relevant ones based on modality:
- Text: bertscore (needs reference), rouge, bleu
- Image: clip_score (no reference needed)
- Audio: utmos (no reference needed)
- Video: temporal_consistency, clip_temporal (no reference needed)
- Output filename? Default to
$ARGUMENTS if provided, otherwise examples/<modality>_eval_custom.yaml
Step 2: Generate the Config
Based on the answers, generate a complete YAML config file. Follow the exact format used in existing examples.
Read an existing example for reference based on the modality:
- Text: read
examples/text_eval.yaml
- Image: read
examples/image_eval.yaml
- Audio: read
examples/audio_eval.yaml
- Video: read
examples/video_eval.yaml
Step 3: Write the rubric
Write a detailed, specific rubric based on what the user said they want to evaluate. Break it into clear dimensions. Each dimension should have a description of what 1 (worst) and 5 (best) mean.
Example rubric format:
Evaluate the response on these criteria:
- Dimension Name: Description. Score 1 if [bad]. Score 5 if [good].
Step 4: Generate test cases
If the user wants generated test cases, create diverse, realistic prompts that cover different aspects of the evaluation. Include:
- A mix of easy and hard cases
- Different topics/styles
- Edge cases where relevant
expected_output for text evaluations (needed for bertscore/rouge)
Step 5: Write and validate
- Write the YAML file using the Write tool
- Run
./judgejudy run <filename> --verbose with --sample 1 to validate it works with a single test case
- If it fails, fix the issue and retry
- Tell the user the file is ready and show them the command to run the full evaluation
Important Rules
- Always use the latest model names (claude-sonnet-4-6, gpt-4o, gemini-2.5-flash, etc.)
- For video generators using wavespeed, include both
model and model_path params
- For audio, include appropriate voice params
- Set reasonable pipeline defaults (concurrency: 3, timeout: 60-300 depending on modality)
- Don't set thresholds unless the user specifically asks for pass/fail criteria
- Keep test case IDs short and descriptive (e.g., "tc-simplify-1", "img-landscape-1")
