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.
An exemplary operational reference in terms of executability and error-recovery workflow, covering generate/edit for images, video (including async Seedance), and speech with concrete commands and validation loops. Its weaknesses are systematic repetition across modality sections, a vague ambient-audio stub, and a monolithic single-file structure that should split per-modality details into reference files.
Suggestions
Split the per-modality details into one-level-deep reference files (e.g. references/video-seedance.md for the two Seedance field tables, capability matrix, and async polling; references/audio-tts.md for the speech API), keeping SKILL.md as an overview with per-modality quick-start examples.
Deduplicate repeated guidance: state the output_dir and session_id field semantics once, and compress the Seedance cost-gate paragraph to a single 'ask the user for resolution before every billable call; never reuse a prior answer' instruction instead of four bolded restatements.
Either make 'Generating ambient audio' actionable (endpoint, request fields, response shape, as done for speech) or remove the section and fold ambient-audio triggers into the speech/TTS section.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly earns its tokens (API specifics Claude cannot know), but there is real padding: the Seedance cost-gate paragraph restates 'ask before every generation' roughly four times in bold, and the output_dir and session_id table rows are repeated verbatim across three request tables. Fits anchor 3 — efficient in places but could be tightened — rather than 4, because the repetition is systematic, not incidental. | 3 / 5 |
Actionability | Nearly everything is copy-paste executable: complete curl commands, request-field tables, success/failure response shapes, per-error_type handling, a ready poll loop, and a real helper script whose documented subcommands match. It stops short of 5 because the 'Generating ambient audio' section is a non-actionable stub ('Generate it programmatically, verify the audio') with no endpoint, fields, or commands. | 4 / 5 |
Workflow Clarity | Clear numbered sequences with explicit validation and feedback loops throughout: verify backend via /api/media/types before any call, magick identify verification for exact sizes, content_filter rephrase-and-retry (up to 3), 400 fallback of dropping aspect_ratio, a full Seedance status-poll loop with hard no-resubmit rules, a 15-minute polling cutoff, and a 6-step last-frame chaining workflow. | 5 / 5 |
Progressive Disclosure | Internal sectioning is clear and the single bundle file (scripts/video_seq.sh) is real and well-signaled, but ~600 lines inlining three distinct sub-APIs (image, Veo video, Seedance video with two field tables and async polling, TTS) in SKILL.md is content that clearly belongs in one-level-deep reference files. Anchor 3 — some structure, but content that should be separate is inline. | 3 / 5 |
Total | 15 / 20 Passed |