Content
77%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.
An exceptionally actionable, safety-conscious reference: every command has executable examples, exact flags, validation checkpoints, and an error-recovery table. Its costs are length and structure — material that belongs in reference files (Agent Intent internals, ring key-management detail) is inlined, and key rules are repeated across the callout, table, and error sections.
Suggestions
Move the Agent Intent deep-dive (enrollment trust model, recovery semantics, keychain/environment requirements) and the ring password-injection/rotation detail into references/ files, keeping in SKILL.md only the commands, required flags, and safety-critical rules.
State each sandbox/TTY/password rule once — the Commands table plus a single callout — and have the Common errors row reference it instead of restating it.
Trim duplicated device-readiness guidance between the top 'Device readiness' callout and the genuine-check row of the error table.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dominated by product-specific rules Claude cannot know (sandbox matrix, provider lists, exit codes), which justifies much of its ~670 lines — but there is real tightening left: sandbox requirements are stated four times (top callout, Commands table column, footnote *, and the USB-timeout error row), device-readiness rules are duplicated between the top callout and the genuine-check error row, and the Agent Intent section carries long prose passages (trust model, relay channel, keychain behavior) that could be condensed or moved to the reference. This matches anchor 3 ('mostly efficient but includes some unnecessary explanation or could be tightened') better than 4, where over-explanation would be only minor. | 3 / 5 |
Actionability | Every flow ships copy-paste-ready commands with exact flags and realistic values — e.g. the WETH wrap/unwrap `--data` examples with padded calldata, the agent-intent status jq polling loops, `echo "destroy" | wallet-cli ring destroy` for non-interactive shells, and a Common errors table mapping each error string to a cause and a fix. Fully executable and covering the common cases, matching the 5 anchor. | 5 / 5 |
Workflow Clarity | Multi-step flows are explicitly sequenced with validation checkpoints and feedback loops: "immediately run `session view`" before asking anything, "Always run with `--dry-run` first to validate calldata before signing", the two-phase Solana withdraw (undelegate → wait ~2–3 days → `--finalize`), the address-mismatch abort-and-genuine-check procedure, fingerprint comparison before approving enrollment, and "confirm with the user before the live run" for irreversible signatures. Destructive operations (send, ring destroy, agent-intent cancel) all carry confirmation/validation guidance, so no cap applies. | 5 / 5 |
Progressive Disclosure | The one bundle reference (references/business-logic.md) exists, is clearly signaled at the top with a purpose statement ("for _why_ a command behaves the way it does… read references/business-logic.md"), and is one level deep — but the bulk of the skill is inlined in SKILL.md rather than split: the full Agent Intent deep-dive (~250 lines: enrollment trust model, recovery semantics, keychain requirements, environment isolation), the ring password-injection and rotation internals, and per-provider swap execution details are all candidates for reference files. This fits anchor 3 ('content that should be separate is inline') better than 4, whose example assumes the bulk lives in the separate file. | 3 / 5 |
Total | 16 / 20 Passed |