Content
92%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: concrete tool-and-command guidance, a verification-first workflow with explicit backup and validation loops, and textbook progressive disclosure that delegates exact paths and failure modes to three real, well-scoped reference files. The only room for improvement is trimming a handful of explanatory justifications in the golden rules.
Suggestions
Trim the rationale clauses in the golden rules (e.g. shorten rule 1 to "Detect the OS and shell first — read get_config's systemInfo and allowedDirectories; paths and commands differ per platform.") to save tokens without losing actionability.
Rule 2's motivation sentence ("The most common reason these sessions dead-end...") could be cut to a single phrase; the instruction to make a tiny test call early is what earns its place.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean, imperative, and free of concept explanations Claude already knows; every section is operational (golden rules, routing, workflow, diagnostic order). It sits at 4 rather than 5 because a few justificatory clauses could be trimmed — e.g. "file paths, commands, and config locations differ across macOS, Windows, and Linux, so confirm the platform before assuming anything" and "The most common reason these sessions dead-end is a provider auth or billing error discovered after a lot of setup work" — which are explanatory padding around an otherwise self-evident instruction. | 4 / 5 |
Actionability | For an instruction-only skill the guidance is fully actionable: every step names the exact Desktop Commander tool to use (get_config for systemInfo, start_process for cp/lsof/curl, edit_block for edits, list_processes for process checks, write_file only from scratch) and includes copy-paste-ready commands like `cp file.json file.json.bak` and `copy` on Windows. It is not at 4 because there are no real gaps — even the Windows variant and post-edit JSON validation are specified. Per the rubric's scoring notes, absence of code blocks in an instruction-only skill is not penalized when guidance is this concrete. | 5 / 5 |
Workflow Clarity | The 6-step general workflow is clearly sequenced with explicit validation checkpoints: "Back up before you edit", "Validate the JSON after every edit by parsing it", "Verify, then stop" with concrete checks (list_processes, lsof/netstat, client reconnected), and "Don't declare success on a guess". Feedback loops exist (validate → fix → re-validate; translate errors and suggest alternatives on failure), and the golden rules function as a checklist — matching the 5 anchor, not 4, because no validation step is left implicit. | 5 / 5 |
Progressive Disclosure | The body is a pure overview that routes to three well-signaled, one-level-deep reference files (references/claude-desktop-mcp.md, references/openclaw.md, references/hermes.md — all verified to exist with substantive content), each with a clear scope description and markdown link, plus official doc URLs for currency-sensitive behavior. Nothing that belongs in a reference is inlined; navigation is trivial. This matches the 5 anchor exactly. | 5 / 5 |
Total | 19 / 20 Passed |