Content
81%Weight 40%Scale 1-5Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body delivers an instruction-only skill with exemplary actionability — exact output format, worked examples for every prompt type, and a fully specified qualify/skip decision procedure with no validation-requiring operations. Its weakness is conciseness: roughly 185 lines where tightened prose (and possibly a reference file for the boundary case-law) could cut a third without losing any decision-relevant signal.
Suggestions
Compress each of the four "user has already told you not to" patterns to its rule plus one short example; several currently spend 4-5 sentences restating the same rationale (e.g., "This wins even when the answer is full of statistics…" duplicates the preceding sentence).
Merge the duplicated educational carve-out material: the "When not to" bullet and the Roth IRA elaboration cover the same ground twice; state the rule once with the single best example.
Consider moving the extended boundary case-law (the four opt-out patterns plus the boundary calls) into a references file, keeping SKILL.md to the core qualify list, prompt-writing rules, and output format.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body runs ~185 lines for a skill whose action is "append 2-3 formatted follow-up questions". The boundary policy is genuinely load-bearing, but several sections restate their own rule: "The user asked you to verify, cite, or flag uncertainty" spends five sentences on one rule, the educational carve-out is explained both in "When not to" and again through the Roth IRA elaboration, and "This wins even when the answer is full of statistics…" repeats the preceding sentence. This matches anchor 3 ("mostly efficient but includes some unnecessary explanation or could be tightened") rather than anchor 4, where only minor trims would be needed. | 3 / 5 |
Actionability | For an instruction-only skill the guidance is fully concrete and copy-paste ready: exact lead-in line ("A few things worth a second look:"), exact format rules ("append it after a blank line", "plain bullets", "No blockquote, no heading", "Plain text only — no HTML, no headings, no emoji", "under ~120 characters", "first person, conversational, question form"), three worked example prompts covering all three prompt types, and a complete example output block. Nothing is left to inference, so this is the anchor-5 case rather than anchor 4's "minor gaps". | 5 / 5 |
Workflow Clarity | The single action is unambiguous and completely sequenced: answer the question fully first ("Always answer the question completely first. The nudge comes after"), apply the enumerated qualify/skip decision including the once-per-conversation rule ("Offer the nudge at most once in a conversation… if you have already offered it on an earlier turn, stay silent"), then append the nudge in the exact format and stop ("The nudge is the closer"). There are no destructive or batch operations, so no validation cap applies, and per the rubric's simple-skill exception a single unambiguous action scores 5; it does not fall to 4 because the decision and output sequence contain no gaps. | 5 / 5 |
Progressive Disclosure | The body is well-sectioned ("Why this exists", "When to offer the nudge", "When not to", "Writing the prompts", "Output format") and appropriately self-contained for a policy the model must apply at reply time — there are no bundle files and no API-reference-style bulk content to split out. However, at ~185 lines of prose with the boundary case-law as the bulk, it exceeds the under-50-lines clean case, and the extended boundary discussion (e.g., the four "user already told you" patterns) could partly live in a reference file. This matches anchor 4 ("good structure; most content is appropriately placed; minor organization gaps") rather than anchor 5. | 4 / 5 |
Total | 17 / 20 Passed |