Content
80%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 well-built instruction skill: lean, executable, and properly factored with a single well-signaled reference file. Its main weakness is the absence of validation feedback loops in a batch workflow — fetch, parse, and generation steps have no checkpoints before changes are staged.
Suggestions
Add a validation checkpoint after the sched.co fetch: verify curl succeeded and the ld+json block was found before parsing, and fall back to manual input (the existing no-URL path) if either fails.
Add a post-creation verification step before 'git add', e.g., confirm each README follows the template's required fields and that no directory slug collides with an existing YYYY/YYYY-MM-DD-event-slug/ directory.
Make Step 1 fully executable by including the extraction command for the ld+json block (e.g., a grep/sed or python one-liner) rather than describing the fields to extract abstractly.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is lean and purely instructional — no concept explanations, just non-obvious operational knowledge (e.g., 'sched.com blocks default agents and WebFetch' with the exact UA header workaround), concrete commands, and terse format rules like 'Location format always includes a flag emoji'. | 5 / 5 |
Actionability | Mostly executable guidance: exact curl/ls/cp/git commands and precise ld+json field mappings ('name → talk title', 'superEvent.name → event name'). It falls short of the top anchor because the ld+json extraction in Step 1 describes what to pull but provides no executable command for it, and the slide 'tokenize... and match' step is direction rather than a runnable recipe. | 4 / 5 |
Workflow Clarity | The sequence (input routing → fetch → create → slides → stage) is clear and includes one checkpoint ('If no match is found, inform the user which talks are missing slides'), but this batch workflow ('Process all URLs in parallel') has no validate-fix-retry checkpoints: silent curl failures, misparsed metadata, or malformed READMEs proceed straight to staging, so it sits at the validation-gaps anchor per the batch-operation cap. | 3 / 5 |
Progressive Disclosure | SKILL.md is a concise overview and the one piece of file-appropriate content (the README layout) is correctly split into a well-signaled, one-level-deep reference: '[readme-template.md](./references/readme-template.md)', which exists in the bundle with exactly the content the body defers to it. | 5 / 5 |
Total | 17 / 20 Passed |