Evals Specify Agent Skill

evals-specify

What this skill does

Conducts bottom-up error analysis following EDD Principles III & IX (Error Analysis & Test Data as Code) to discover and document draft evaluation criteria from human observation of system failures.

Output:

• Draft Eval Records - Individual EVAL-*.md files in .adlc/drafts/evals/ with open coding notes
• Error Pattern Documentation - Bottom-up failure taxonomy from actual traces
• Pass/Fail Examples - Real examples that should pass/fail each criterion
• Auto-handoff to /evals-clarify for axial coding and clustering

Key EDD Principles Applied:

• Principle III: Error Analysis & Pattern Discovery - Open coding → failure taxonomy
• Principle IX: Test Data as Code - Dataset planning and coverage analysis
• Principle II: Binary Pass/Fail - Maintain strict binary pass/fail conditions
• Principle V: Trajectory Observability - Track full multi-turn conversation traces

When to use

• Starting evaluation development: No existing criteria, need discovery from failure logs
• Production incident analysis: Recent failures require systematic analysis
• Quality assessment: Discovering and codifying boundary conditions from failures

When NOT to use

• No failure traces/specs: Generate synthetic traces first, or use /evals-init to set up security baselines
• Known criteria already exist: Use /evals-clarify to refine or /evals-implement to generate code

Process

User Input

$ARGUMENTS

Treat user input as specific failure areas or error patterns to analyze (e.g., "authentication bypass", "RAG irrelevant results").

• --traces N — Number of traces to analyze (default: 20, min for theoretical saturation)
• --source SOURCE — Trace source location (e.g., logs, support tickets)

Execution Steps

Phase 1: Open Coding Analysis

• Reviews the user-provided failure logs or spec requirements.
• Conducts open coding of traces to discover recurring failure patterns (EDD Principle III).
• Identifies: core problem, causal conditions, and consequences.

Phase 2: Create Draft Criteria

Group patterns into draft criteria. For each:

• Define strict Pass Condition (observable, binary yes/no)
• Define strict Fail Condition (observable, binary yes/no)
• Document real pass/fail examples directly from traces

Phase 3: Create Draft Files

• Copy skills/evals/evals-templates/eval-criterion-template.md to .adlc/drafts/evals/EVAL-{NNN}.md.
• Populate metadata and error analysis notes.
• Regenerate index at .adlc/drafts/evals/evals.md.

Phase 4: Auto-Handoff

Trigger /evals-clarify for axial coding and clustering.

Verification

• Draft files created at .adlc/drafts/evals/EVAL-*.md
• Index file .adlc/drafts/evals/evals.md updated with draft summaries
• Each draft contains: status "draft", pass/fail conditions, trace sources, and concrete examples
• Auto-handoff context produced with list of created drafts