Content
81%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 highly actionable and operationally safe: complete executable snippets, explicit validation and error-recovery checkpoints, and precise semantics for every operation and result shape. Its main weakness is repetition — the reply-immediately/verify-after-recovery rules are stated three separate times, which costs tokens without adding clarity.
Suggestions
State the "reply immediately, call no more tools, verify with get_sandbox_info after recovery" rule once (e.g., in a single shared section) and reference it from the upgrade and restart sections instead of repeating it in the quoted outputs, the per-section prose, and Safety and persistence.
Trim or shorten the full quoted result.content blocks for upgrade and restart to the operationally relevant lines, since the following prose paragraphs already paraphrase the same information.
Consolidate the overlapping bullets in "Safety and persistence" (lines about resumption, cancellation, and no-auto-retry largely restate the restart and upgrade sections) into a short checklist that only adds new information.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The post-recovery guidance ("Reply to the user now and do not call more tools" / "use get_sandbox_info to verify") appears in the quoted tool outputs and is then restated nearly verbatim in the prose after both the upgrade and restart sections, and a third time in "Safety and persistence". Sections are otherwise code-first and efficient, but this triple repetition is a clear tightening opportunity, fitting "mostly efficient but could be tightened" better than the minor-trimming level of 4. | 3 / 5 |
Actionability | Three complete, copy-paste-ready run_sdk_snippet blocks with imports, a consistent error pattern (raise SystemExit(1) after printing result.content), branch handling on data["operation"], and concrete result.content/result.data examples for every outcome (including the already_current case). Fully executable and covering the common cases. | 5 / 5 |
Workflow Clarity | Despite involving destructive operations, every checkpoint is explicit: user-intent gates before upgrade/restart, saving work to .workspace before a rebuild, the immediate-reply rule, post-recovery verification via get_sandbox_info, and a no-auto-retry error loop ("Inspect the returned error, and after recovery use get_sandbox_info before deciding what to do next"). This clears the destructive-operation cap and matches the explicit-validation-with-error-recovery anchor. | 5 / 5 |
Progressive Disclosure | No bundle files exist, and the ~175-line body is cleanly sectioned (terminology, result reading, one section per operation, safety) so navigation is easy. However, it exceeds the under-50-line simple-skill case, and the duplicated post-recovery prose and long typical-output blocks leave minor organization gaps — good structure, but not the ideally split/consolidated layout of a 5. | 4 / 5 |
Total | 17 / 20 Passed |