View on GitHub →

Model Prompt Tests

Structured prompts for comparing frontier model behavior

  • Markdown
  • 5-Point Rubric Scoring

A Markdown library of prompts for evaluating models across coding, planning, writing, and reasoning - each with a scoring rubric.

Problem

Model evaluations shouldn’t be vibes. How do you systematically compare how models handle planning, coding constraints, risk detection, and instruction-following?

Approach

Each prompt is a self-contained test case with a rubric scored 1-5. The prompts cover:

Coding & Engineering: Build a Pomodoro timer, debug JavaScript debounce, senior PR review, architecture database choice.

Planning & Advice: Build a six-hour SQL learning plan, plan a San Francisco city break.

Writing & Creativity: Rewrite an internal announcement, generate app names under constraints.

Reasoning & Evaluation: Diagnose business metrics, follow five-bullet format, handle ambiguous requests, detect safety risks in production migration scenarios, evaluate evaluation rubrics.

The scoring rubric focuses on planning quality, judgment calls, instruction adherence, technical accuracy, clarification-seeking behavior, and practical usefulness - not just whether the final answer looks polished.

Outcome

A personal eval harness for judging model quality beyond benchmarks. Run the same prompt against multiple models with consistent system instructions, score with the rubric, compare behavior.