Content
75%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 highly actionable, well-sequenced operational skill whose commands are copy-paste ready and whose workflow includes validation and destructive-op safeguards. Its weakest point is token efficiency: several dense prose policy blocks and setup notes could be trimmed or moved to reference files.
Suggestions
Move the full 'Environment contract' and 'Preparation failures' policy prose into the referenced compute-env-contract.md (or a references/ file) and keep only a 2-3 line digest inline, cutting roughly 15 lines of the densest text.
Trim the trailing setup blockquotes (Vast.ai/Modal/W&B setup) to bare commands and links, removing explanatory asides like 'ideal for users without a local GPU who need to debug code or run small-scale tests'.
Add a short explicit error-recovery branch after Step 5 (e.g. what to check if screen -ls shows no session or the log shows an immediate crash) to close the last workflow-clarity gap.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dominated by executable commands and project-specific conventions Claude would not know, but the dense prose blocks ("Environment contract", "Preparation failures") and trailing setup blockquotes (e.g. "ideal for users without a local GPU who need to debug code") are convoluted or padded and could be tightened. | 3 / 5 |
Actionability | Fully executable, copy-paste-ready commands cover the common cases for every environment: complete rsync include/exclude rules, ssh+screen launch with conda activation, W&B logging snippets, Modal launcher flow, and the full vastai rent-to-destroy lifecycle. | 5 / 5 |
Workflow Clarity | The 7-step workflow is clearly sequenced with explicit skip conditions, a pre-flight GPU check, a verify-launch step, smoke-test validation, and careful handling of the destructive instance-destroy operation; it falls short of the top anchor only because explicit error-recovery branches for launch failures are not spelled out (the retry/escalation policy is prose). | 4 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ absent), so all guidance is inline in one 340-line file; the numbered step sections are well-organized and sibling-skill delegation is clearly signaled, but the referenced ../shared-references/compute-env-contract.md is outside the bundle and its policy content is partially duplicated inline — good structure with minor organization gaps. | 4 / 5 |
Total | 16 / 20 Passed |