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.
This is a strong, actionable debugging runbook that supplies genuinely non-obvious domain knowledge with concrete code, commands, and a sequenced workflow including validation feedback loops. The main weakness is mild redundancy across the Ground Rules, Artifact Hygiene, and Isolate sections, and the absence of any bundle files means progressive disclosure rests entirely on in-section organization.
Suggestions
Consolidate the recurring artifact-hygiene and do-not-commit reminders into a single section to remove redundancy and tighten conciseness.
Add a small executable tensor-compare helper snippet (np/torch vs mx with atol/rtol and first-N-elements printout) so the compare step is fully copy-paste ready.
Consider extracting the detailed inject-initial-latents procedure into a references file and summarizing it in SKILL.md to improve progressive disclosure for this longer-than-50-line skill.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with novel domain knowledge (RNG parity gotchas, MPS NaN behavior, scheduler indexing, the inject-initial-latents trick) and avoids explaining basics Claude already knows, though artifact-hygiene reminders recur across several sections and could be tightened. | 4 / 5 |
Actionability | Provides executable code (the randn_tensor export snippet), specific commands (uv run python, HF_HUB_OFFLINE=1, /usr/bin/time -p), concrete file layout, and tolerance values, with only minor gaps (e.g., no full copy-paste compare script and some placeholder names). | 4 / 5 |
Workflow Clarity | The default workflow is clearly sequenced (pick repro, decide checkpoints, export, compare) with a binary-search feedback loop and pass/fail validation, but validation is not framed as a hard gate before proceeding, keeping it just below a 5. | 4 / 5 |
Progressive Disclosure | No bundle files exist; the single SKILL.md is well-organized into clearly headed sections with only one-level references to sibling skills (mflux-model-porting, mflux-testing) and no nested reference chains, though the long Isolate procedure could conceivably be split into a reference file. | 4 / 5 |
Total | 16 / 20 Passed |