evals-skills

Generate a custom trace annotation web app for open coding during LLM error analysis. Use when the user wants to review LLM traces, annotate failures with freeform comments, and do first-pass qualitative labeling (open coding). Also use when the user mentions "annotate traces", "trace review tool", "open coding tool", "label traces", "build an annotation interface", "review LLM outputs", or wants to manually inspect pipeline traces before building a failure taxonomy. This skill produces a tailored Python web application using FastHTML, TailwindCSS, and HTMX.

Build a structured taxonomy of failure modes from open-coded trace annotations. Use this skill whenever the user has freeform annotations from reviewing LLM traces and wants to cluster them into a coherent, non-overlapping set of binary failure categories (axial coding). Also use when the user mentions "failure modes", "error taxonomy", "axial coding", "cluster annotations", "categorize errors", "failure analysis", or wants to go from raw observation notes to structured evaluation criteria. This skill covers the full pipeline: grouping open codes, defining failure modes, re-labeling traces, and quantifying error rates.

Use this skill when crafting, reviewing, or improving prompts for LLM pipelines — including task prompts, system prompts, and LLM-as-Judge prompts. Triggers include: requests to write or refine a prompt, diagnose why an LLM produces inconsistent or incorrect outputs, bridge the gap between intent and model behavior, reduce ambiguity in instructions, add few-shot examples, structure complex prompts, or improve output formatting. Also use when the user needs help distinguishing specification failures (unclear instructions) from generalization failures (model limitations), or when iterating on prompts based on observed failure modes. Do NOT use for general coding tasks, document creation, or non-LLM writing.

Build, validate, and deploy LLM-as-Judge evaluators for automated quality assessment of LLM pipeline outputs. Use this skill whenever the user wants to: create an automated evaluator for subjective or nuanced failure modes, write a judge prompt for Pass/Fail assessment, split labeled data for judge development, measure judge alignment (TPR/TNR), estimate true success rates with bias correction, or set up CI evaluation pipelines. Also trigger when the user mentions "judge prompt", "automated eval", "LLM evaluator", "grading prompt", "alignment metrics", "true positive rate", or wants to move from manual trace review to automated evaluation. This skill covers the full lifecycle: prompt design → data splitting → iterative refinement → success rate estimation.