Skip to content

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

5 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Eval Coach

An Agent Skill for designing comprehensive AI evaluation strategies using Evaluation-Driven Development (EDD).

About

Created by Bayram Annakov while building Onsa.ai - AI agents for B2B sales prospecting.

If you find this useful, say hi on LinkedIn!

What is Eval Coach?

Eval Coach guides you through a structured 5-step framework for evaluating LLM applications:

  1. Define Success - Map business goals to measurable metrics
  2. Design Dataset - Create diverse test cases (happy path, edge cases, adversarial)
  3. Select Methods - Choose Automated, LLM-as-Judge, or Human evaluation
  4. Plan Automation - Integrate evals into CI/CD
  5. Monitor Production - Track drift and collect feedback

Installation

Claude Code / Cursor / VS Code

Copy this skill to your project's skills directory:

git clone https://github.com/BayramAnnakov/eval-coach.git ~/.claude/skills/eval-coach

Or add to your project:

mkdir -p skills
git clone https://github.com/BayramAnnakov/eval-coach.git skills/eval-coach

Usage

Invoke the skill by name or with trigger keywords:

/eval-coach

Or just mention evaluation-related topics:

  • "help me create an evaluation strategy"
  • "design test cases for my agent"
  • "set up LLM testing"

The 50-40-10 Rule

A practical distribution for evaluation methods:

Tier Method Cost Percentage
1 Automated (schema, keywords, latency) $0.00/run 50%
2 LLM-as-Judge (quality, relevance) $0.01-0.05/run 40%
3 Human Review $5-50/run 10%

Templates Included

  • templates/dataset.py - LangSmith dataset creation with example test cases
  • templates/evaluators.py - 10 ready-to-use evaluators (automated, LLM-as-Judge, performance)
  • templates/compare.py - Experiment comparison utilities

Key Insight: Silent Failures

The most dangerous failures are the ones your tests don't catch. The input_data_consistency_evaluator was born from a real production issue:

We ran an agent with --target john_smith --company CompanyA, but John actually works at CompanyB. The agent silently reconciled the contradiction by writing "engaged with CompanyA via partnerships" instead of flagging the mismatch. All 9 evaluators passed, but the output was misleading.

Always include adversarial test cases that probe for silent failures.

Agent Skill Standard

This skill follows the Agent Skills open standard (SKILL.md format), supported by:

  • Claude Code
  • Cursor
  • VS Code with AI extensions
  • Gemini CLI
  • OpenAI Codex

Requirements

For the templates:

  • Python 3.9+
  • LangSmith account (for dataset management)
  • Google API key (for LLM-as-Judge evaluators)
pip install langsmith langchain-google-genai

Configuration

Set the judge model via environment variable:

export GOOGLE_API_KEY="your-api-key"
export JUDGE_MODEL="gemini-3-flash-preview"  # Default (fast, cost-effective)
export JUDGE_MODEL="gemini-3-pro-preview"    # For higher quality judging

License

MIT License - see LICENSE

Related Resources

About

Agent Skill for Evaluation-Driven Development (EDD) - guide AI evaluation strategies with the 50-40-10 rule

Topics

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages