Content
88%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 well-executed process skill: fully executable commands, a specified mapping format, explicit validate-and-rerun feedback loops around the batch reconciliation run, and a concrete output contract. Its minor costs are a trimmable motivational opening and an inline report template/helper reference that could be split into a bundled reference file.
Suggestions
Trim the two opening paragraphs to one or two sentences — the operational stance ('evidence, not rulings', never edit the originals) is load-bearing, but the narrative about drifting legacy processes can be compressed.
Move the step-7 report template and the step-4 helper-function reference into a bundled reference file (e.g. references/report-template.md) and link to them, keeping SKILL.md as the workflow overview.
State file-extension triggers (.pdf, .xlsx) in the description to lift trigger term coverage to the top anchor.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with non-obvious operational guidance ("A wrong mapping looks like a disagreement too", "Never describe a sampled result as 'every input'", owner questions must "Never mention scripts, fields, cells or formulas") and assumes Claude's competence throughout. Not 5: the two opening paragraphs (~90 words of motivation — "They drift apart, and nobody notices because each person only uses one of them") and a few framing sentences ("This is the one step that needs judgment about meaning") could be trimmed without losing operational content; not 3 because padding is minor and localized rather than 'some unnecessary explanation' spread across sections. | 4 / 5 |
Actionability | Copy-paste-ready commands for each phase (`uv run python inspect_sources.py path/to/form.pdf path/to/calculator.xlsx [SHEET]`, `uv run python reconcile.py --mapping mappings/<form>.py --out runs/<form>/results.json`), a fully specified mapping-file shape (PDF, XLSX, SHEET, INPUTS, NAMES, CELLS, OUTPUTS, OUTPUT_LABELS), named helper functions with signatures (`load(mapping)`, `effective_weights(values, scores, ...)`, `conditions.explain(rows, mask, ...)`, `owner_questions.translate(co, text, terms)`), and a complete report template. Specific examples cover the common cases via the worked Field Visit example; not 4 because there is no pseudocode or missing key detail — every instruction maps to an executable command or file. | 5 / 5 |
Workflow Clarity | Steps 0–7 are clearly sequenced (triage → inventory → mapping → exhaustive check → facts → guide claims → owner questions → report), and the batch operation has an explicit validation feedback loop: "Check the mapping before trusting the results. Run a small sample first... Correct the mapping and rerun until every remaining disagreement is explained by the logic, not by the mapping", plus a defined failure mode ("reconcile.py stops with an assertion... fix the mapping, don't edit the artifacts") and a coverage-sampling rule when combinations exceed a few million. Not 4: validation checkpoints are explicit with error-recovery loops, not merely present. | 5 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/ directories), and the body is well-sectioned with numbered headers and clearly signaled repo-tooling references (`reconcile.py`, `inspect_sources.py`, `facts.py`, `owner_questions.py`, `mappings/field_visit_risk.py`) that are one level deep. Not 5: the inline report template (~25 lines) and the helper-function reference list are content that could live in a reference file, and at ~125 lines the skill exceeds the 'under 50 lines, no external references needed' case that the rubric exempts; not 3 because structure is good and nothing is buried or nested. | 4 / 5 |
Total | 18 / 20 Passed |