How to Build a Custom AI Model Benchmark Using Claude Code

Based on our work together, can you help me brainstorm how I AI benchmark and eval set we can test every time a new model comes out to consistently score different tasks That would be relevant to our podcast audience.

Let's actually focus on task for builders, PRDs prototypes, agentic, multi-step and agentic voice basically does a pass the vibe check in my open claw.

Now, build the complete evaluation harness for the tasks we've defined. Generate the following components:
1. A test runner script that can execute the same prompts across multiple models, like [Opus 4.8, GPT-5.5, and Sonnet 5].
2. A backend scorer that uses an LLM-as-judge to evaluate the outputs based on a clear rubric.
3. A local HTML page that displays the outputs from all models side-by-side for a blind human review, allowing me to enter a 1-5 score and notes for each.

Confirm the generated harness includes both an automated LLM-as-judge scorer and an interface for manual human scoring, or a 'vibe check'. This hybrid system allows you to combine objective performance metrics with your own subjective, taste-based feedback.

My actual values:
[insert the files, settings, accounts, or constraints for this step]

Give me the exact commands, settings, or output to use. Finish with a pass or fail check for implement a hybrid scoring system.