Content
72%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 compact, readable skill that clearly explains its three modes and honestly documents its advisory enforcement limits. Its main weakness is workflow rigor for a destructive-operation skill: Freeze mode lacks any verification step and the post-confirmation/denial flow is undefined, which caps workflow clarity. The Integration section also references ECC-specific tooling the skill never defines.
Suggestions
Add an explicit verification step for Freeze mode, e.g., "Before every Write/Edit, verify the target path is inside the frozen directory; if not, stop and tell the user what outside change is needed."
Define the confirmation flow end-to-end with an example prompt (what the agent shows, what counts as confirmation, and what to do on denial — e.g., propose the safer alternative instead).
Trim or explain the Integration section: "codex -a never" and "ECC 2.0 observability risk scoring" reference tooling the skill never defines, so either expand them into actionable guidance or remove them.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly lean, but there are minor trim opportunities: Guard mode's "The agent can read anything but only writes to src/api/, and asks before any destructive command" restates the two combined modes, "Useful when you want an agent to focus on one area" pads Freeze mode, and the Integration section's ECC-specific items ("Enable by default for `codex -a never` sessions", "Pair with observability risk scoring in ECC 2.0") reference tooling undefined in the skill. This fits the score-4 anchor (efficient, minor instances of over-explanation) rather than 5, where every token would earn its place. | 4 / 5 |
Actionability | Guidance is mostly concrete for an instruction-only skill: a specific watched-pattern list ("rm -rf", "git push --force", "DROP TABLE", "docker system prune", etc.) and concrete behaviors ("The agent does not Write/Edit outside src/components/, and explains why when a change would need it"). Minor gaps remain — no example of the confirmation prompt the agent should show, and "Unlock: Tell the agent to stop applying the rules" leaves the mechanism vague — so it fits score 4 rather than the fully copy-paste-ready score 5, and is more complete than the score-3 pseudocode level. | 4 / 5 |
Workflow Clarity | Careful mode has an explicit checkpoint ("shows what the command does, asks for confirmation, suggests safer alternative"), but this is a destructive-operation skill and validation is incomplete: Freeze mode has no verification step (e.g., checking the target path is inside the frozen directory before each write), and there is no defined flow for what the agent does after the user confirms or denies. Per the guideline capping destructive-operation skills with validation gaps at 3, this is 3 rather than 4. | 3 / 5 |
Progressive Disclosure | The ~55-line body has well-organized sections (When to Use, How It Works with one subsection per mode, Enforcement) and no bundle files, none of which are needed at this size. The two external links (permission deny rules, PreToolUse hooks) are appropriate one-level-deep pointers. Per the simple-skill exception for short skills with no need for external references, well-organized sections alone warrant 5; score 4 would imply misplaced content, of which there is none. | 5 / 5 |
Total | 16 / 20 Passed |