Content
75%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 dense, expert-level instruction skill with concrete slot patterns, per-mode mistake/repair pairs, an explicit compression order, and multiple validation gates before delivery. The main costs are repetition of the Director's Read gate, checklist rules compressed into hard-to-scan run-on prose, and a workflow spread across many sections instead of one canonical ordered view.
Suggestions
State the Director's Read load-and-classify gate once (in Director Formula) and reference it from the opening and Prompt Build Process instead of repeating the full instruction three times.
Reformat the eight-point check in the Pre-Delivery Screen as a short bulleted checklist (or move it entirely into direct-for-the-model.md and link to it) rather than one multi-clause sentence, so each rule is scannable at delivery time.
Consolidate the process into a single numbered end-to-end sequence (classify → lane record → reference roles → shot table → draft → checks → screen → output contract), with Mode Gate, Time Structure, and Compression Rules as clearly-labeled substems, so the reader does not have to infer the order across sections.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body carries almost no filler and assumes Claude's competence — no section explains what video generation or prompting is — with all content being skill-specific rules. It loses a point to redundancy (the Director's Read load-and-classify instruction is stated three times: lines 23, 35, and 79; multishot-grammar is loaded in both the line-27 sweep and Time Structure) and to overlong chained sentences ('Before the screen, run the eight-point check...' packs an entire checklist into one paragraph) that could be split without adding tokens. | 4 / 5 |
Actionability | Concrete guidance dominates: the director-formula table gives prompt-ready example phrases per slot ('slow dolly-in from medium product shot to macro label detail'), the mode table names each mode's common mistake and its literal repair ('Say `preserve @Image1 exactly`'), compression is an explicit priority-ordered cut list, and the Output Contract is a numbered deliverable list. It falls short of fully executable because the eight-point check and Stakes and Peak rules are compressed into prose rather than a runnable checklist, and no complete worked example prompt appears inline (examples are delegated to examples-by-mode.md). | 4 / 5 |
Workflow Clarity | The Prompt Build Process section gives a real ordered sequence ('First, load the Director's Read... Next, assign reference roles... then write the draft... Finally, run a self-check') backed by explicit validation checkpoints (self-check, directing coherence test, paper render on the shot table, moderation pre-screen, anti-slop pass). It is not a 5 because the full workflow is distributed across six sections the reader must assemble, some ordering is stated inversely ('Before the screen, run the eight-point check'), and intermediate checkpoints like the paper render are invoked without a link to where they are defined. | 4 / 5 |
Progressive Disclosure | The body is a genuine overview that delegates depth to ~12 one-hop reference files, each with a clear purpose signal ('for the checklist', 'for image-to-video', 'for shot order'), plus routing to sibling skills for continuation and filtering — navigation is easy. It stays at 4 because the load-sweep in the opening references paragraph bundles eleven files into one sentence, and substantial reference content (the eight-point check from direct-for-the-model, the Stakes and Peak summary from directing-engine) is duplicated inline rather than split out, and no bundle files were present alongside this SKILL.md to verify the referenced paths exist. | 4 / 5 |
Total | 16 / 20 Passed |