Content
92%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 an unusually rigorous operating manual: executable commands for every operation, hard validation gates with named error-recovery fallbacks at each risky step, and a clean one-level split of library-edit recipes into references. The only weakness is density — a few rationale passages and a duplicated ST_Xstring note could be trimmed without losing any operational content.
Suggestions
State the ST_Xstring escaping behavior once (either in "Verify before reporting success" or "Limits") and cross-reference it from the other section; it is currently explained in full in both places.
Trim the rationale for the 2 ms freshness tolerance (the double-rounding of epoch milliseconds, why a wider window hides edits) down to the actionable rule: compare size and mtimeMs with a 2 ms tolerance, and on mismatch ask the user to re-select.
Consider moving the detailed xlsx anchor-comparison recipe (how to run the containment test on formatted cells — joining text cells with spaces, excluding formatted cells) into a references file, keeping only the rule and the pointer in SKILL.md.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Nearly every line is non-inferable operational knowledge — refusal rules (shared-formula groups, non-allow-listed docx elements), the 2 ms freshness tolerance, number-format asymmetries — so it never explains what Claude already knows. It is not a 5 because a few passages are justification rather than instruction (the multi-sentence rationale for why the double rounding of epoch milliseconds yields a 2 ms tolerance) and the ST_Xstring escaping note is explained in full twice, in both "Verify before reporting success" and "Limits". | 4 / 5 |
Actionability | Fully executable throughout: complete copy-paste commands for both scripts (uv run --with openpyxl python scripts/office_extract.py --file ... --anchor ... --out ...), exact anchor and edit JSON shapes in tables, a ready one-liner for reading whole-millisecond mtimes, and a runnable verification snippet. All referenced paths (scripts/office_extract.py, scripts/office_patch_copy.py, references/docx-edit.md, references/pptx-edit.md) exist in the bundle. | 5 / 5 |
Workflow Clarity | Every risky step carries an explicit validation gate — the freshness check ("the anchor check below is not optional"), the anchor containment test with stop-and-ask-user failure handling, the expectText hard gate, staged writes that leave nothing behind on failure, and a dedicated "Verify before reporting success" section ("If verification fails, say so and show the error"). Not a 4 because there are no validation gaps, and each refusal names its fallback route (e.g. "Rewrite such a range with openpyxl", "see 'Edit docx' below"), giving feedback loops for every failure mode. | 5 / 5 |
Progressive Disclosure | The body holds the universal procedure and appropriately splits the per-library edit recipes one level deep into references/docx-edit.md and references/pptx-edit.md, each clearly signaled at the exact decision point ("Read [references/docx-edit.md] before writing any python-docx edit"); the scripts are bundled and their paths resolved. Not a 4 because the split is clean, references are one level deep, and nothing that belongs in a separate file is buried inline. | 5 / 5 |
Total | 19 / 20 Passed |