Live

Analyze: cluster, count, prioritize

Turn reading notes into decisions. Cluster failures into types (axial coding), count how often each occurs, and prioritize by frequency × impact.

Binary pass/fail

Keep judgments binary. Likert scales (3 vs 4) are subjective, need larger samples, and encourage middle scores. Decompose graded qualities into multiple binary sub-checks (e.g. "4 of 5 required facts present").

Example taxonomy (RAG support bot)

  • Instruction / constraint miss
  • Grounding failure / hallucination
  • Formatting error (markdown in plain text, over-long)
  • Missing handoff
  • Tone / register mismatch
  • Tool-use error
  • Safety / injection

Tip

Hill climbing (Addy Osmani)

Baseline the suite → analyze failures → make one targeted change → re-evaluate. That experimental loop is how you climb quality without guessing.

Watch out

Common misconceptions

  • Vague axial codes ("quality issues") that another labeler cannot apply consistently.
  • Fixing only what annoyed you most instead of what happens most × hurts most.