AI Prompt Reference
Log in
Back to library
DevelopmentGPT-5WorkflowTestingData

AI Feature Evals Suite

Design offline evals and human review criteria for AI features.

AI Feature Evals Suite

Prompt

# Role
You are an AI product evaluation engineer familiar with offline evals, human labeling, and regression tests.

# Task
Design an evaluation suite for the AI feature below.

# Inputs
Ask me to provide or paste: product/project background, target audience, use case, constraints, existing assets, desired language, and channel. If data, screenshot descriptions, competitors, or past examples are included, prioritize those facts and do not invent missing details.

# Output Format
1. Task definition
2. Test sample stratification
3. Scoring rubric
4. Failure taxonomy
5. Baseline and thresholds
6. Human review flow

# Quality Bar
1. Samples must cover edge cases
2. Metrics relate to user experience
3. Avoid only subjective quality

# Avoid
Do not provide generic advice. Do not use unverifiable hype. Do not pad the answer with irrelevant completeness. If business, medical, financial, legal, or security risks are involved, clearly state assumptions and boundaries.

# Process
First decide whether the information is sufficient. If key information is missing, ask up to five clarifying questions. If enough information is available, produce the actionable version directly. Then provide three optional improvement directions for iteration.

# Final Deliverable
End with a concise copy-ready version that preserves the key constraints, output structure, and quality bar.

Curated by the editorial team · Updated 06/28/2026 · Model: GPT-5

Usage guide

How to use this prompt

This template is designed for development tasks. Replace the sample details with real constraints before running it in GPT-5.

  1. Step 1

    State the stack, runtime, inputs, outputs, and existing constraints.

  2. Step 2

    Ask for the approach and risks before requesting the smallest verifiable change.

  3. Step 3

    Run tests, type checks, and critical scenarios locally before merging.

Details to replace or add

Specific inputs produce more useful results. Do not submit passwords, private information, or confidential business data.

  • Language, framework, and versions
  • Current code and error output
  • Expected inputs and outputs
  • Compatibility and performance constraints
  • Acceptance test cases

Output checklist

  • The code runs on the specified versions
  • Edge cases and errors are handled
  • Existing project patterns are reused
  • Tests cover critical behavior
  • No new security or performance risk appears

Common adjustments

Provide the directory structure and interfaces when the answer drifts from the project.

Limit files and request staged changes when the proposal is too broad.

Ask for runnable test commands and expected output when verification is unclear.

This prompt separates Development, Workflow, Testing, Data requirements into context, constraints, and output format. Keep the objective fixed and revise only the conditions that failed before rewriting the whole template.

Related prompts

Explore more templates in Development.