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 well-engineered instruction skill: routing table up front, executable memsearch commands throughout, explicit human-in-the-loop validation at every irreversible step, and a clean split of platform-specific detail into verified one-level-deep references. The only deductions are minor: a repeated platform-reference pointer that could be stated once, and section C's transcript-drill command being delegated to another skill's reference instead of being written out.
Suggestions
In section C, inline at least one concrete form of the transcript-drill command (or a canonical example invocation) instead of deferring entirely to the memory-recall skill's platform reference, so the workflow is executable without a second lookup.
State the platform-reference pointer once (e.g., in the existing reference list near the top) and refer to it by short name from sections B, D, and Install paths, rather than repeating 'see your platform reference file' three times.
Compress section A's 'Be exact — do not guess' prose into a short checklist (verify against transcript / keep step general / never fabricate) to save tokens without losing the instruction.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and imperative with almost no generic explanation of concepts Claude already knows — everything is MemSearch-specific. Minor trimmable fat remains: 'see your platform reference file' is repeated in sections B, D, and Install paths, and the 'Be exact — do not guess' paragraph in A is wordier than a compact checklist would be. Not 5: a few sentences could be tightened without losing meaning; not 3: there is no padded or redundant explanation block. | 4 / 5 |
Actionability | Nearly all guidance is copy-paste executable: 'printf ... | memsearch skills add --name ... --body-file -', 'memsearch skills status', 'git -C .memsearch/skill-candidates log --oneline -5', and the config get/set commands with defaults shown. The one real gap is section C's key step: 'Run the transcript drill (see the memory-recall skill's platform reference for the exact command)' delegates the exact command rather than stating it. Not 5: a central command of workflow C is not written out; not 3: every other workflow has complete, concrete commands with flags and expected outputs explained. | 4 / 5 |
Workflow Clarity | Stages are explicitly labeled (0→1→2) with intent routing up front, and each flow carries explicit validation checkpoints and feedback loops: 'Before recommending or installing, skim the candidate's body... this is the last chance to catch a wrong step', 'skills status' hash comparison to detect drift, the interactive install checkpoint with confirmed destination, and the empty-list fallback offering A/C/D. Not 4: verification is explicit and repeated at every irreversible step (install, config set), plus a Guardrails checklist. | 5 / 5 |
Progressive Disclosure | The body is a ~135-line overview and all five bundle files (references/claude-code.md, codex.md, openclaw.md, opencode.md, dsh.md) exist and are cited by exact relative path at the point of need, one level deep. Platform-specific config prefixes and install paths are correctly split out of the main body with clear signaling. Not 4: every reference is well-signaled, shallow, and verified against the actual bundle structure. | 5 / 5 |
Total | 18 / 20 Passed |