Content
65%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 a well-organized, token-efficient set of behavioral rules that mostly assumes Claude's competence, and its single-file structure is appropriate for its length. Its weaknesses are operational vagueness in several directives (verification cadence, memory implementation, the referenced 'send_to_user' tool) and the absence of any sequenced workflow despite governing long-horizon, multi-step runs.
Suggestions
Make the vague directives executable: replace "establish a method for checking your own work at a specific interval" with a concrete recipe (e.g. 'every N completed steps, spawn a fresh-context subagent with the spec and the diff and ask it to list discrepancies'), and specify the memory file path/format in section 5.
Add a short ordered run workflow (e.g. 1. restate the goal and scope, 2. plan and delegate, 3. execute with interval verification, 4. final fresh-context audit, 5. outcome-first report) so the protocol sequences its policies for long-horizon runs.
Either name the actual client-side tool interface for mid-task updates (section 7's 'send_to_user') with how to invoke it, or drop the reference to a tool that may not exist in the target environment.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean, directive bullet prose with no padding and no explanations of concepts Claude already knows; e.g. "Do not stop mid-task to ask for permission for reversible actions." A few bullets are abstract and could be trimmed ("Being readable and being concise are different things, and readability matters more"), keeping it at anchor 4 rather than 5. | 4 / 5 |
Actionability | Some directives are concretely operational ("Pause for user input ONLY for: (1) destructive/irreversible actions ...", "Present the clarification as a definitive set of choices (e.g., Option A vs. Option B)"), but many others are abstract exhortation without operational detail — "establish a method for checking your own work at a specific interval" and "work end-to-end with high reliability" give no executable specifics. This lands on anchor 3 (some concrete guidance but incomplete) rather than 4 (mostly executable guidance). | 3 / 5 |
Workflow Clarity | The eight sections are organized policies rather than a sequenced workflow: there is no ordered run procedure (e.g. plan → execute → verify → report) for a protocol claiming to govern multi-day runs. Section 3 does provide verification guidance ("Verify your work against the specification, preferably using fresh-context subagents", "Report outcomes faithfully"), so checkpoints exist but are implicit and unsequenced — anchor 3, not 4 (which requires a clear sequence) and not 2 (validation is not absent). | 3 / 5 |
Progressive Disclosure | The skill is a single self-contained ~57-line SKILL.md with no references/, scripts/, or assets/ directories; nothing in the body needs to live in a separate file. Eight clearly headed sections make navigation easy, matching the simple-skill exception (well-organized sections, no external references needed) for a top score. | 5 / 5 |
Total | 15 / 20 Passed |