Content
52%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 SKILL.md body is well-structured as a gated, per-operation index with clear sequencing and real consent/validation checkpoints, but the bundle does not deliver what it promises: every referenced rules/ and templates/ file is missing, so the skill cannot actually be executed past phase 0 of any operation. Redundant restatement of the gates across three sections further inflates the token cost of an index whose substance lives elsewhere.
Suggestions
Ship the referenced bundle files: all 11 rules/*.md files and templates/pointer-snippet.md linked from the 'Required Reading by Operation' table are absent from the bundle, dead-ending every operation at phase 0. Add them, or inline the minimal content each operation needs to run.
Collapse the triple restatement of the write gates (Core Workflow gate column, Core Principles, and Definition of Done) into a single location, and link each rules/*.md file once per section instead of repeating it in every row of a phase table.
Link references/research-sources.md directly (e.g., from the Reflexion note in Core Principles) and remove its dangling ../rules/scaling-tiers.md pointer, so the one bundle file that exists is actually reachable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body avoids teaching known concepts and uses compact tables, but the same content is restated three times: Core Principles ("Two-phase write... Extract candidates first, then resolve each..."), Anti-patterns ("Re-recording the same fact instead of UPDATE-ing"), and Definition of Done ("Candidate list shown to the user with ADD / UPDATE / DELETE / NOOP tags") all re-encode the phase-table gates, and phase tables repeat the same rules/*.md link up to five times in one table. This is beyond anchor 4's 'minor instances that could be trimmed' but short of anchor 2's padded prose, so 3 fits best. | 3 / 5 |
Actionability | Concrete elements are present (exact paths like `~/.agent-memory/<scope>/`, named files INDEX.md/AUDIT.log, `$0 == "write"` parsing, a literal pointer snippet, a literal summary string "Saved N memories to <scope> (<tier>)."), but every phase's operational substance — the never-store list contents, candidate extraction, merge criteria, staleness cutoffs — is deferred to rules/*.md files that do not exist in the bundle, so key details are missing rather than minor. That places it at anchor 3 ('some concrete guidance but incomplete'), not 4. | 3 / 5 |
Workflow Clarity | Each operation has an explicitly sequenced phase table with a named gate per phase plus a global gate rule ("Do not proceed to the next phase until the prior phase's gate passes"), and destructive operations carry real validation ("User saw the entries and explicitly confirmed", consent preview, AUDIT.log), so the destructive-operation cap does not apply. Not 5: no error-recovery feedback loops are specified for failed gates, and the step-level detail lives in the missing rules files, leaving minor validation gaps at anchor 4. | 4 / 5 |
Progressive Disclosure | The written design matches the top anchor (thin index, per-operation 'Required Reading' table, one-level-deep links), but scoring against the actual bundle shows all 11 referenced rules/*.md files and templates/pointer-snippet.md are absent — every on-demand read dead-ends. The only bundle file present, references/research-sources.md, is never directly linked from SKILL.md and itself points to a missing ../rules/scaling-tiers.md. The disclosure structure is claimed but not delivered, which is worse than anchor 3's 'references present but not clearly signaled'. | 2 / 5 |
Total | 12 / 20 Passed |