Read the operating agreement and analysis procedure and artifacts. For local work, use local review guidance; for an existing managed workflow, use managed guidance.
For bundled local tools and exact commands, read toolkit setup. Run the workspace installer for this skill; it preserves existing customized files. For the managed route, use concrete service operations after confirming the connected schema.
Start from reviewed evidence
Locate existing annotations, their source traces, sampling rules, and any taxonomy revisions. Preserve original notes and author/status information. Inspect the behavior behind the notes before grouping it. If there are no human observations, prepare a review queue and use eval-discover if installed; agent-generated observations remain suggestions.
Reuse a managed platform's existing analysis and exports when available. For either local track, perform the analysis over local records and extend the existing review UI. No grader framework or new hosted connection is required.
Propose and refine
Group notes by the observable failure and expected behavior. Use semantic search or embeddings only to find candidate neighbors; inspect evidence to decide membership. Similar wording can describe different failures, and different wording can describe the same failure.
Give each proposed mode a stable ID, clear definition, inclusion/exclusion rules, representative evidence, and boundary examples. Allow multiple modes per annotation and an unresolved queue. Separate observed symptoms from hypothesized root causes. Do not force every record into a category or turn a broad label such as “hallucination” into a supposedly actionable taxonomy.
Present groups beside original annotations and traces. Let the reviewer merge, split, rename, move, reject, and confirm proposals. Preserve revision history and reviewer decisions. Human approval of a category does not automatically confirm every assignment or produce per-criterion Pass/Fail labels. If review is unavailable, finish the proposed taxonomy, evidence links, and review queue without claiming confirmation.
Prioritize and hand over
Count distinct cases per mode and state the reviewed denominator and sampling purpose. Multi-label counts can overlap. Rank using observed frequency, product impact, and uncertainty; rare severe failures must remain visible. Do not infer production prevalence from a discovery sample.
Save a versioned taxonomy, annotation-to-mode mappings, unresolved items, and a concise analysis report using the reference's local artifact convention or an equivalent existing format. Verify evidence links and save/reload/export behavior when changing a UI. Report unverified interactions honestly.
Hand confirmed modes and examples to case curation and grader design (eval-dataset and eval-grade when installed). With a trusted baseline, use the prioritized development failures to guide eval-descent. Keep held-out evidence out of the optimization loop.