Content
67%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 tightly written, checkpoint-rich instruction body with concrete gates, refusal scripts, and a crisp output contract. Its critical defect is that the entire execution apparatus — seven reference files and one script — is referenced but absent from the bundle, and via "../../references/" paths that escape the skill directory, leaving the skill non-self-contained and its navigation broken.
Suggestions
Ship the referenced bundle files (directors-read.md, continuation-handoff.md, sequence-project-state.md, prompt-compiler.md, reference-transfer-contract.md, continuity-qc.md, failure-atlas.md, directing-engine.md, and scripts/extract_last_frame.py) inside the skill directory, or inline the essential record/compile steps so SKILL.md is self-contained.
Fix the reference paths to live within the skill folder (references/directors-read.md, not ../../references/directors-read.md) so links resolve relative to SKILL.md.
Assemble the scattered steps (load read, classify, gate, observe, compile, output) into one ordered workflow section so the sequence and its validation checkpoints are readable end-to-end without jumping between sections.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense domain-specific rule-making (input gate fields, canon rule, scene boundary rule, extension-depth cap) with no generic concept explanations Claude already knows. Minor trimmable material keeps it from 5: the philosophical "Intent" paragraph ("The soul of this skill is fidelity to what actually happened..."), the duplicated Director's-Read load instruction at the top and in the reference list, and the "never paste internal labels into final generation prose" rule stated twice. Not 3 because there is no genuinely unnecessary explanation, just slight tightening opportunities. | 4 / 5 |
Actionability | Concrete, executable guidance throughout: an exact refusal script to quote verbatim, a runnable command ("python scripts/extract_last_frame.py <take>"), a 15-item required-input checklist, five defined continuation types, and a numbered 10-item output contract. Not 5 because the core compilation steps ("complete its canonical narrative or non-narrative record", "compile the exact visible or audible carriers") are delegated to reference files that are not present in the bundle, so a reader cannot fully execute from this file alone; not 3 because what is written here is specific and directly executable rather than pseudocode. | 4 / 5 |
Workflow Clarity | A real sequence with explicit checkpoints: source gate before any prompt is written, observation confidence flags ("observation_confidence: low", "requires_user_confirmation: true") when pixels aren't in hand, the intent echo "confirmed before generation spends money", the extension-depth re-anchor cap, and the scene-boundary rule. Not 5 because the end-to-end order of operations is never assembled into one flow — it is distributed across the intro, gate, and rules sections — and error recovery is delegated to absent references (continuity-qc, failure-atlas); not 3 because checkpoints are explicit and well-placed, not merely implied. | 4 / 5 |
Progressive Disclosure | References are clearly signaled with load conditions ("Load failure-atlas... when the continuation failed or drift is visible"), but the bundle contains only SKILL.md — none of the seven referenced files (directors-read.md, continuation-handoff.md, sequence-project-state.md, prompt-compiler.md, reference-transfer-contract.md, continuity-qc.md, failure-atlas.md, directing-engine.md) nor scripts/extract_last_frame.py exist, and the links point to "../../references/", two levels above the skill directory and outside it. Every load instruction leads to a missing file, so the overview is an orphaned index with broken navigation. Not 3 because the signaling quality cannot compensate for the fact that the referenced structure simply is not provided; not 1 because the SKILL.md itself is well-sectioned and the references are one conceptual level deep. | 2 / 5 |
Total | 14 / 20 Passed |