Content
87%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
The body is concise and highly actionable with specific model IDs, CLI commands, and a curl fallback, organized into clean sections appropriate for a sub-50-line skill; its only gap is the absence of a verification checkpoint for the batch image-generation step.
Suggestions
Add an explicit validation step before returning asset ids, e.g. poll each request with 'muapi predict wait <request_id>' and confirm success, retrying failed keyframes, to satisfy the batch-operation feedback-loop expectation.
Note how to handle partial failures (some keyframes fail) so the agent does not silently return a broken sequence.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean — an inputs table, four numbered steps, short notes — and assumes Claude's competence without explaining what keyframes or story arcs are, fitting the 'lean and efficient; every token earns its place' anchor. | 3 / 3 |
Actionability | It gives concrete, executable specifics: model id 'nano-banana-2', aspect_ratio=16:9, explicit prompt construction, CLI commands ('muapi image generate', 'muapi predict wait'), and a full curl fallback — copy-paste-ready guidance rather than vague direction. | 3 / 3 |
Workflow Clarity | Steps are clearly sequenced (decompose → per-beat generate → run parallel → return ids), but this is a batch operation with no validation that images succeeded before returning asset ids, so per the rubric's batch-operation feedback-loop cap it stays at 2 rather than 3. | 2 / 3 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are absent) and the skill is under 50 lines with well-organized sections (Inputs, Steps, Notes, Trigger Keywords), meeting the simple-skill allowance for a 3 with no external references needed. | 3 / 3 |
Total | 11 / 12 Passed |