Content
77%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 is exceptionally actionable and the workflow is superbly sequenced with validation checkpoints and recovery paths. Its weaknesses are token efficiency — heavy duplication of the scope-limits block and repeated rules — and monolithic inlining of material (test specs, protocols, templates) that belongs in reference files for progressive disclosure.
Suggestions
Replace the two verbatim 20-line SCOPE LIMITS blocks with a single pointer to ../shared-references/review-scope-limits.md (the file the skill already references for nightmare mode), saving ~40 lines.
Move the four Acquittal Gate test specifications and the Debate Protocol templates into a references/ file (e.g. references/acquittal-gate-tests.md), keeping only the operative rules in SKILL.md.
Consolidate the stop-condition and reviewer-memory rules, each currently stated in 3–4 places, into one authoritative statement with cross-references.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~500-line body duplicates the entire 20-line SCOPE LIMITS block verbatim in both the medium-mode prompt and the round-2+ template (despite a shared-references file already existing for it), restates the stop condition in four places, and repeats memory/debate rules across the dedicated section and Phases B.5/B.6. Mostly efficient with no conceptual padding, but clearly could be tightened — anchor 3. | 3 / 5 |
Actionability | Fully executable guidance throughout: exact spawn_agent/send_input blocks with model and reasoning effort pinned, complete JSON state schema, JSONL receipt format with field-level rules, run_id generation recipe, concrete render command with flags, and explicit prioritization rules for fixes. Copy-paste ready — anchor 5. | 5 / 5 |
Workflow Clarity | The multi-phase workflow is clearly sequenced (Initialization decision tree covering fresh/resume/stale/legacy cases, Phases A–E, stop gate B.5.1 deliberately placed after the memory update), with explicit validation checkpoints, feedback loops (debate protocol, regression-triggered trace diffing), and four test specs validating the acquittal gate. Anchor 5. | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent) and all referenced shared-references live outside the skill, unverifiable here. The body is well-sectioned but monolithic: the four Acquittal Gate test specifications, the full debate/memory protocols, and the duplicated prompt templates clearly belong in separate reference files — anchor 3 ('content that should be separate is inline'). | 3 / 5 |
Total | 16 / 20 Passed |