tavva/eval-design

Eval Design

Guide users through designing production-quality LLM evaluations. Output: structured spec a coding agent can implement with Langfuse.

Announce at start: "I'm using the eval-design skill to help design your evaluation."

When to Use

digraph when_to_use {
  "User question" [shape=box];
  "About measuring LLM output quality?" [shape=diamond];
  "Use this skill" [shape=box, style=filled, fillcolor=lightgreen];
  "Not this skill" [shape=box];

  "User question" -> "About measuring LLM output quality?";
  "About measuring LLM output quality?" -> "Use this skill" [label="yes"];
  "About measuring LLM output quality?" -> "Not this skill" [label="no"];
}

Use for:

• Designing new evaluations
• Choosing between eval approaches
• Creating scoring rubrics or judge prompts
• Building golden datasets
• Questions like "how do I know if my LLM is working?"

Not for:

• Langfuse setup, tracing, or observability (use the official langfuse skill)
• Implementing evals (output spec to coding agent)
• Generic testing that isn't LLM-specific

The Process

digraph eval_design_flow {
  rankdir=TB;

  understand [label="Understand\nthe System" shape=box];
  failures [label="Identify\nFailure Modes" shape=box];
  match [label="Match Eval Type\nto Problem" shape=box];
  design [label="Design\nthe Eval" shape=box];
  output [label="Output\nSpec" shape=box];

  understand -> failures -> match -> design -> output;
}

Ask questions ONE AT A TIME. Adapt depth based on user's experience level.

Phase 1: Understand the System

Ask about:

• What does the LLM application do? (summarisation, Q&A, agent, chat)
• What's already instrumented in Langfuse? (traces, spans, generations)
• What does a "good" output look like for users?

Phase 2: Identify Failure Modes

Key questions:

• What failures have you seen in production?
• Have you done error analysis on real traces?
• Which failures have the highest business impact?

Critical: Ground eval design in ACTUAL failures, not hypothetical concerns. If user hasn't analysed real failures, recommend starting there.

Phase 3: Match Eval Type to Problem

| Problem Type | Recommended Eval | |--------------|------------------| | Format validation (JSON, dates) | Code-based | | Factual accuracy with known answers | Code-based + match | | Tone, helpfulness, coherence | LLM-as-judge | | Safety, toxicity | LLM-as-judge | | RAG retrieval quality | LLM-as-judge per chunk | | High-stakes decisions | Human → LLM-judge | | Novel/unclear failure modes | Human annotation first |

Consult references/eval-types.md for detailed guidance on each type.

Phase 4: Design the Eval

For each eval, define:

1. Success criteria - prefer binary PASS/FAIL over Likert scales
2. Rubric or logic - specific, unambiguous criteria
3. Dataset requirements - happy path + edge cases + adversarial
4. Integration - when/where to run

Use references/judge-prompts.md for LLM-as-judge prompt templates.

Output Format

Output a structured spec using the template in references/document-template.md.

The spec should be complete enough that a coding agent can implement it using the Langfuse SDK without further clarification.

Principles for Great Evals

Start from Error Analysis

Examine real failures, not hypothetical concerns. Generic metrics ("hallucination score") without grounding in actual user complaints are noise.

Questions to surface:

• "What would a human expert check when reviewing this output?"
• "If this eval passes but users are unhappy, what did we miss?"

Match Tool to Task

Don't over-engineer:

• Code-based evals are faster, cheaper, more deterministic
• Only use LLM-as-judge when judgment is genuinely subjective
• Human annotation for calibration, not production eval

Binary PASS/FAIL Over Likert

Rating 1-5 creates noise and inconsistency. Binary forces clarity:

• What exactly makes something PASS?
• What exactly makes something FAIL?

Cover the Right Cases

Dataset should include:

• Happy path: Common, expected inputs
• Edge cases: Boundaries, unusual but valid inputs
• Adversarial: Inputs designed to break the system

Calibrate LLM Judges

Before deploying LLM-as-judge:

1. Create human-labelled examples (20-30)
2. Run judge on same examples
3. Measure agreement (target >85%)
4. Adjust prompt until calibrated

Red Flags to Watch For

Push back when you see:

• Jumping to LLM-as-judge when code-based would suffice
• Likert scales instead of binary
• Generic metrics without connection to actual failures
• Missing adversarial cases
• No plan for calibration

Anti-Patterns

"Vibes-based" evaluation: Testing a few examples and shipping if it "looks good" → Systematic dataset with coverage of failure modes

Evaluating the model, not the product: Generic benchmarks that don't reflect real use → Evals grounded in actual user complaints and failures

Over-engineering: Complex eval pipelines for simple checks → Simplest eval that catches the failure mode

Handoff to Implementation

After outputting the spec:

"This eval design is ready for implementation. A coding agent can use the official langfuse skill to implement this using the Langfuse SDK."

Do NOT attempt to implement the eval. The skill outputs design specs only.

Quick Reference

| User Says | Do This | |-----------|---------| | "How do I know if my LLM is working?" | Start Phase 1 | | "I need to evaluate [specific thing]" | Jump to Phase 2, focus on that failure mode | | "Should I use LLM-as-judge?" | Ask about failure mode → recommend based on type | | "Help me write a judge prompt" | Use references/judge-prompts.md templates | | "What should my dataset look like?" | Cover happy path + edge + adversarial |

Additional Resources

Reference Files

• references/document-template.md - Output spec template with guidance
• references/eval-types.md - Detailed eval type selection and patterns
• references/judge-prompts.md - Ready-to-use LLM-as-judge prompts