Content
70%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 highly actionable with a cleanly sequenced, well-validated five-step workflow that leans correctly on bundled writer scripts. Its weaknesses are verbosity — heavy repetition and very long paragraphs that re-state the same constraints — and weak progressive disclosure, with large posture/template/failure-mode material inlined rather than moved into reference files.
Suggestions
Consolidate the repeated 'materialized once' and 'never detached / foreground on this thread' rules into a single authoritative statement in 'How you speak while it runs' or 'How to behave', and have other sections reference it rather than restating it.
Break the 300–700-word paragraphs (Purpose, flow step 3's pull/transform, flow step 5's join) into shorter bullet-or-step units so the workflow reads as a checklist rather than prose.
Move the per-platform gate templates (lines 79–94), the Expected-audience-size posture (lines 157–165), and the Failure-modes catalog (lines 183–196) into separate reference files (e.g. references/gate_templates.md, references/audience_size.md, references/failure_modes.md) and link to them from the main flow.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~196-line body is markedly verbose with several padded/repetitive sections: 'materialized once' recurs across Purpose, Works with, and the flow steps; the 'never detached / foreground' rule is restated in Purpose, How you speak, step 3, How to behave, and Failure modes; and 'the script is the source of truth, not memory' repeats verbatim, with many 300–700-word single paragraphs. It stays accurate and on-topic (not the 'severely padded' score-1 case) but the repetition and prose density clearly exceed the 'minor instances' of score-4, landing at the noticeably-verbose score-2 anchor. | 2 / 5 |
Actionability | Fully executable guidance throughout: exact shell commands with full paths and flags ('--list-identifiers', '--contact-window', '--preflight', '--download', '--phase transform', '--phase join'), precise tool-call shapes ('entity_find' with 'expression_string', 'max_identifiers: 3', 'format: csv', 'audience_limit', 'offset'), the download-driver JSON contract ('"done": <bool>, "expired": <bool>', re-run until done), and copy-paste-ready confirmation/gate examples — matching the score-5 anchor. | 5 / 5 |
Workflow Clarity | The five-step flow (Preflight → Pick platforms/confirm → Run export → Deliver → Add device IDs) is clearly sequenced with explicit validation checkpoints — a liveness probe, a disk preflight before the pull, entity-cursor coverage gating, and re-run-until-'done' feedback loops for downloads and transforms — plus halt-and-surface rules, matching the score-5 anchor; the destructive/batch cap-at-3 does not apply because validation is present throughout. | 5 / 5 |
Progressive Disclosure | The real bundle (scripts/writers/{_common,meta,google,reddit,tiktok}.py and audience_size_range.py) is correctly used and clearly signaled one level deep, with deterministic math deferred to the scripts — but no reference/asset files exist at all, and the gate templates, the audience-size posture, and the full Failure-modes catalog are inlined into SKILL.md as a near-monolithic prose wall rather than split into separate files, fitting the score-3 anchor 'content that should be separate is inline' while staying above score-2 because script separation is handled well. | 3 / 5 |
Total | 15 / 20 Passed |