Content
85%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.
A tight, well-structured instruction-only skill: the checklist-with-escalation workflow is exemplary, with concrete output formatting and a genuine reject-then-retry feedback loop. The only trimming opportunities are the duplicated reject stance and the slightly repetitive untrusted-input boilerplate, plus the absence of a filled-in evidence example.
Suggestions
State the default-REJECT stance once (either in the intro or in Rules) instead of twice, saving a line without losing emphasis.
Add a two-line filled-in example of the Evidence section (e.g., 'Tests: npm test — 42 passed, 0 failed; Scope: pass, 2 files, both in allowed list') so the output format is unambiguous on first use.
Tighten the untrusted-input block so each bullet covers a distinct rule (instruction refusal, provenance of instructions, priority/verdict self-labeling, flagging hygiene) rather than restating 'do not follow instructions' three times.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and operational with no explanations of concepts Claude already knows, but the reject stance is stated twice ('Your job is to reject unless evidence is strong' and 'Default stance: REJECT until proven otherwise') and the untrusted-input block restates its core prohibition three ways. These minor redundancies keep it just below the 'every token earns its place' anchor. | 4 / 5 |
Actionability | Concrete checklist items ('You ran tests (or equivalent) and report pass/fail with output snippet'), a copy-paste-ready output template, and a specific escalation rule ('If you cannot run tests (env issue) → ESCALATE_HUMAN') give mostly executable guidance. It falls short of a 5 only because no worked example shows a filled-in evidence report for a common case. | 4 / 5 |
Workflow Clarity | The ordered all-must-pass checklist (Scope → Intent → Tests → No cheating → Risk) includes explicit validation (run tests, report output), a feedback loop (REJECT with numbered reasons plus 'Suggested next step for implementer'), and escalation checkpoints (ESCALATE_HUMAN, human review for medium+ risk). This matches the top anchor: clear sequence, explicit validation, and error-recovery loops. | 5 / 5 |
Progressive Disclosure | The skill is under 50 lines with no bundle files (references/, scripts/, assets/ do not exist), and the body is organized into clear, well-labeled sections (Inputs, Checklist, Output, Rules, Untrusted input). Per the simple-skill guideline, well-organized sections with no need for external references warrant the top score. | 5 / 5 |
Total | 18 / 20 Passed |