Content
77%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 disciplined, agent-native workflow document: imperative, specific, and unusually strong on validation with its start/completion gates, pre-apply revalidation, and review verdicts. The weaknesses are structural rather than instructional — reference material is inlined in a monolithic file, and a small amount of constraint repetition could be consolidated.
Suggestions
Move stable reference material into one-level-deep files (e.g. references/plan-shape.md for the plan template and decision taxonomy, references/target-state.md for the status.json/fork-ledger schemas, references/potion-example.md for the file-by-file inspection list) and keep SKILL.md as the mode/dispatch overview.
Make verification concrete: name the actual commands per package manager (e.g. `pnpm typecheck`, `pnpm test --filter`) or state how to derive them, so apply-mode verification is copy-paste ready instead of a placeholder.
Consolidate the duplicated whole-file-overwrite rules (stated in Core Take, Three-Way Classification, and Apply Mode) and the planning-mode constraints (repeated between Flow Modes and Planning Mode) into single authoritative sections.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense, imperative, and assumes Claude's competence throughout — no space is spent explaining what Plate, registries, or three-way merges are. It falls short of anchor 5 due to mild redundancy: planning/apply constraints appear in both "Flow Modes" and the dedicated "Planning Mode"/"Apply Mode" sections, and the whole-file-overwrite caution is stated three times ("Core Take", "Three-Way Classification", "Apply Mode"). | 4 / 5 |
Actionability | Guidance is overwhelmingly concrete: exact paths (`apps/www/src/registry/registry-editor.ts`, `<target>/.plate-ui-sync/runs/<date>-<scope>/plan.md`), a full `status.json` schema, invocation examples (`sync-plate-ui plan ../potion code-block`), a copyable apply payload, plan-shape and output templates. Not anchor 5 because a few steps stay abstract — "run focused target verification" and "run focused target typecheck, lint and test" name no commands, and verification commands are left as `<commands/results>` placeholders. | 4 / 5 |
Workflow Clarity | The multi-step process is explicitly sequenced with validation checkpoints: a "Start Gates" checklist, mode-by-mode may/must-not rules, a pre-apply loop (re-read plan, recompute classification, confirm row still valid, record verification commands), "Completion Gates" for both planning and apply, and a review mode with verdicts (fresh/stale-plate/stale-target/blocked) that functions as an error-recovery feedback loop. This is an anchor-5 match — checklists plus feedback loops for a destructive-capable workflow — and the destructive-operation cap does not apply since validation is present, not missing. | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are all absent), so the entire ~500-line workflow lives inline in SKILL.md. Section headers make it navigable, but self-contained reference material — the `status.json` schema, the Potion file-by-file inspection list, the plan-shape and dashboard-payload templates — is exactly the content that belongs in one-level-deep reference files. That matches anchor 3 ("some structure... content that should be separate is inline"); it is above anchor 2 because the inline material is well organized, and below anchor 4 because nothing is split out. | 3 / 5 |
Total | 16 / 20 Passed |