Content
81%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 exceptionally actionable and well-validated workflow document — exact commands, flags, paths, a hard fidelity gate, and per-failure-mode fixes. Its main weakness is token efficiency: the engine rule and several safety rules are each repeated many times, inflating context cost without adding information, and it carries one dangling `tests/` reference.
Suggestions
State the 'ALWAYS GPT Image 2, HTML is only a finishing step' rule once in Decision Rules and trim its re-statements from the Purpose, Inputs (`route_hint`), Phase 1, Phase 2A heading, and Spend Reference — this alone would cut significant length without losing information.
Consolidate the repeated colour discipline ('never invent a colour') and text-stacking rules into single entries in Decision Rules / Failure Modes instead of restating them in Brand grounding, Quality Checks, and Failure Modes.
Remove or fix the `tests/` reference (the directory is absent from the bundle), and consider moving the Quality Checks detail and Spend Reference into a reference file so SKILL.md stays a lean overview.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with genuinely non-obvious operational detail, but the same rules are repeated many times: "ALWAYS GPT Image 2 / never HTML as the generator" appears in the Purpose, Inputs (`route_hint`), Phase 1, the Phase 2A heading, Decision Rules, Failure Modes, and Spend Reference; "never invent a colour" and "never stack two text layers" are each stated 3+ times. This is beyond the "minor instances" of anchor 4 — the document could lose a third of its length by stating each rule once in Decision Rules. | 3 / 5 |
Actionability | Guidance is copy-paste concrete throughout: the exact renderer command (`node <goose-graphics>/screenshot/screenshot.js --format <canvas> --input index.html --output render.png --font-delay 1500`), exact FAL slug and flags (`--aspect_ratio 3:4 --quality high --resolution 2k`, upscale via `fal-ai/esrgan`), concrete file paths (`scripts/cutout_product.py`, `assets/overlay-template.html`), aspect→canvas mappings, and a per-slot copy-authoring procedure. It fully specifies what to do for both generation paths and the common failure cases. | 5 / 5 |
Workflow Clarity | The phases (0 → 0.5 → 1 → 2A/2B → 3) are clearly sequenced with an explicit validation checkpoint — the "fidelity gate (3 checks, all must pass)" — plus a Quality Checks section with concrete pass/fail criteria and feedback loops (re-roll on the original reference, overlay text for text-only failures, reject for non-remixable references) and a Failure Modes section pairing each cause with a fix. | 5 / 5 |
Progressive Disclosure | Structure is good: operational detail sits in clearly-signaled bundle files that actually exist (`scripts/cutout_product.py`, `assets/overlay-template.html`), external capabilities are referenced one level deep by path, and sections are well-organized with headers. Not anchor 5: the body is a ~230-line monolith where QC detail, failure modes, and spend tables are all inline, and the referenced `tests/` directory does not exist in the bundle — a dangling reference. | 4 / 5 |
Total | 17 / 20 Passed |