Content
85%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
A well-structured, actionable extraction skill with a clear gated workflow and appropriate one-level-deep reference disclosure. Its main weakness is redundancy in the intake/clarification section, where option lists and gating sentences are repeated.
Suggestions
Collapse the duplicated clarification option lists (bare invocation vs. broad request) into one parameterized list to remove repeated bullets and the repeated 'Do not search transcript evidence or draft a quote set until the user selects one' sentence.
Consolidate the threshold guidance scattered across steps 2 and 4 (relevance 0.7 -> 0.6, 20-40 expansion, confidence >= 0.8, theme relevance down to ~0.65) into a single compact thresholds reference to reduce repetition.
Trim the repeated scope/filter reminders in step 1 (the account/segment bounded-pass guidance restates rules already covered by the clarification gate).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body assumes Claude's intelligence (no basic concept explanations) but carries redundancy: the clarification option lists are repeated for the bare-invocation and broad-request cases, and the gating sentence 'Do not search transcript evidence or draft a quote set until the user selects one' appears multiple times, so it could be tightened. | 2 / 3 |
Actionability | Provides highly concrete, executable guidance for an instruction-only skill — explicit thresholds (customer_confidence >= 0.8, theme_relevance >= 0.75), named fields (speaker_role_guess, customer_confidence), search sizing (10-15 then 20-40), a copy-ready Markdown output template, and default counts (five quotes, one per call). | 3 / 3 |
Workflow Clarity | A clearly numbered five-step sequence (Resolve scope, Search, Extract, Verify/rank, Deduplicate) with explicit gating checkpoints ('Do not search... until the user selects one', 'report 0/[target] when none pass threshold') and sparse-results retry logic providing a feedback loop. | 3 / 3 |
Progressive Disclosure | SKILL.md acts as the overview and signals a single one-level-deep reference — [references/extraction-and-output.md] — for the detailed extraction/ranking/JSON rules, and that referenced file exists in the bundle, matching the well-signaled one-level-deep anchor. | 3 / 3 |
Total | 11 / 12 Passed |