Content
61%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 packs genuinely expert, non-generic keyword-research methodology into a small token budget, with concrete examples, labeling rules, and fallbacks. Its weaknesses are structural: no headers or sections turn it into a wall of text, the workflow sequence and checkpoints stay implicit, and a few passages could be tightened.
Suggestions
Break the nine paragraphs into section headers (e.g., '## Seed and boundaries', '## Variant probing', '## Broader queries', '## Competitor and SERP evidence', '## Handoff and stopping'), which would address both progressive_disclosure and the implicit sequencing.
Make the workflow order explicit — a short numbered sequence for the discovery pass, with the labeling rules (observed vs. inferred, hypotheses) presented as checkpoints per step.
Move the worked domain examples (background remover, song generator) into a short '## Examples' section or a reference file so they are clearly signaled rather than interleaved mid-paragraph.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with domain-specific guidance — variant-probing dimensions, verbatim examples like the "generate song with ai" expansion, labeling rules (observed vs. inferred), and recording fields — with almost no explanation of concepts Claude already knows. It falls short of 5 because several tangled sentences ('For a narrow seed, identify the constraint that narrows it... Show the parent followed by the input seed and explain the difference') and hedging qualifiers ('when relevant', 'when helpful') could be tightened. | 4 / 5 |
Actionability | For an instruction-only research skill the guidance is concrete: exact surfaces to probe (Autocomplete, People Also Ask, People Also Search For, Related, GSC queries, ad search terms), exact fields to record (URL, market, language, date, source), a worked variant example, labeling rules, and a fallback procedure via an authorized cloud browser. It is not a 5 because some guidance stays abstract — the dimension lists ('Explore alternative wording, action, input, desired output, audience...') tell Claude what to vary but not how to systematically generate or prioritize variants. | 4 / 5 |
Workflow Clarity | A rough sequence exists in paragraph order (seed and boundaries → base expression family → broader/parent queries → competitor pages → SERP surfaces → first-party data → clustering handoff), and there are good implicit checkpoints like 'Keep unattested forms labeled as hypotheses' and 'Distinguish a module absent on an inspected page from a page that could not be inspected'. This matches anchor 3: the sequence is present but steps and checkpoints are implicit rather than clearly staged, with no explicit ordering or stop-conditions-per-step. | 3 / 5 |
Progressive Disclosure | The body is a title followed by nine unbroken prose paragraphs with no section headers, subheadings, or lists, matching 'minimal structure' (anchor 2). While the ~30-line skill needs no external reference files, it is not 'well-organized sections' as the simple-skill exception requires, and embedded worked examples (the background-remover and song-generator cases) read as candidate reference material. | 2 / 5 |
Total | 13 / 20 Passed |