Content
78%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.
The body is a dense, highly actionable rulebook: concrete commands, executable snippets, and real named examples cover every phase from collection design to deployment and recovery, with validation steps in the right places. Its weaknesses are structural rather than substantive — a 425-line monolith with no bundle references, and a workflow whose order must be inferred from section arrangement rather than an explicit sequence.
Suggestions
Split stable, detailed material — the UI client/Result contract and the Git recovery procedures — into files under references/ and link them from short SKILL.md sections, turning the monolith into an overview with one-level-deep references.
Add a brief ordered overview near the top (inspect an existing app → define collections → write actions → build UI → verify in a temp home → deploy only when asked) so the phase sequence is explicit rather than assembled from section headers.
Move the per-field edge-case rules (readOnly semantics, JSON Merge Patch null handling, read-tolerance of `other`/`problems`) into a collections reference file, keeping only the design decisions and one example per rule inline.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The prose is ruthlessly compressed and project-specific ("Write last.", "Annotate honestly.", "Never keep two writable copies in step") with no explanation of concepts Claude already knows, fitting anchor 4's 'efficient; minor instances that could be trimmed'. It is not a 5 because the 25 KB monolithic body includes sections (UI client contract details, recovery minutiae) whose volume could be reduced by offloading to reference files. | 4 / 5 |
Actionability | Guidance is fully executable throughout: a copy-paste-ready temp-home verification shell block, complete TypeScript snippets (`await words.update({ path }, (word) => ({ reps: (word?.reps ?? 0) + 1 }))`, `connect<typeof tasksApp>()`), exact commands (`epicenter new --from`, `git restore --source=<commit> -- apps/<app>/data/`), and named real examples (Tasks' `add`, Vocab's `review`) covering the common cases. This matches anchor 5 rather than 4, whose 'minor gaps' are not evident. | 5 / 5 |
Workflow Clarity | Clear phased structure (decide files → actions → UI → verify → server → recovery) with genuine validation points ("run `epicenter build <app>` to check it before opening its page", "Read each record's `other` and `problems` afterward", the failed-build feedback loop), matching anchor 4. It falls short of anchor 5 because the end-to-end sequence is never laid out as an ordered path — the reader must assemble the order from section headers and scattered rules, and some checkpoints (e.g., typecheck placement) are implicit. | 4 / 5 |
Progressive Disclosure | The bundle contains no reference, script, or asset files at all; everything, including the full UI client contract and Git recovery procedures, is inlined in a ~425-line SKILL.md. Section headers are well organized and external pointers are clearly named (`packages/app/README.md`, `../product-collapse/SKILL.md`, `query-layer`), which keeps it above anchor 2, but content that clearly belongs in separate reference files is inline — the anchor-3 pattern. | 3 / 5 |
Total | 16 / 20 Passed |