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.
A tightly engineered routing-and-invariants body: lean, assumption-free of Claude's knowledge, concrete commands and verbatim strings where they matter, a numbered freshness ladder with stop conditions, and an exemplary need→reference table where every referenced file exists in the bundle. The consistent 4s reflect redundancy across sections, run-on compression that burdens parsing, and validation expressed as rules rather than ordered check-then-proceed loops.
Suggestions
Consolidate the repeated simulate-first / read-scaffolding-before-init rules (Answer Contract, Boundary and Preflight, Operational Invariants 2, 9) into one canonical 'Simulation preflight' statement the other sections reference, trimming tokens and contradiction risk.
Break the 150–250-word single-sentence rules (Answer Contract, Operational Invariant 5, Artifact Fit rule 6) into short enumerated clauses so each clause is independently checkable.
Give at least the main path an explicit ordered flow (init → scaffold → simulate with the literal command → deploy preflight) with check-then-proceed validation steps, instead of expressing the same gates as scattered prohibitions.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and lean with essentially zero explanation of concepts Claude already knows — every line is a rule or a pointer. It falls short of 5 only through redundancy: simulate-before-deploy and 'load project-scaffolding/simulation.md first' instructions recur across the Answer Contract, Boundary and Preflight, and Operational Invariants 2 and 9, and the run-on sentences (e.g., the ~200-word Answer Contract sentence, Artifact Fit rule 6) could be split without adding tokens. | 4 / 5 |
Actionability | Mostly executable guidance: literal commands like 'cre workflow simulate <workflow-dir> --target <target-name> ...', 'cre whoami', 'cre init', 'cre login', exact URLs, the exact TypeScript secret idiom 'runtime.getSecret({ id }).result().value', and the verbatim confirmation string 'I can file this issue for you. Reply `file it` to confirm.'. Minor gap: the simulate invariant demands 'concrete values' while the template itself carries placeholders, and bulk detail (config shapes, handler code) is deferred to references rather than shown. | 4 / 5 |
Workflow Clarity | Sequences are clear where given — the Freshness section is a numbered 1–8 fetch ladder with explicit checkpoints ('Stop when the page answers. Fetch at most three pages', 'Ask for the missing fact only'), and risky operations are gated with validation (simulate before deploy, testnet writes require preflight approval plus immediate second confirmations, feedback drafts require explicit confirmation before filing). Not a 5: there is no single end-to-end ordered flow; the body is a rule collection whose ordering a reader must reconstruct, and some checkpoints are stated as prohibitions rather than check-then-proceed loops. | 4 / 5 |
Progressive Disclosure | The 'Need → Read' routing table cleanly splits the skill across all 15 references (each named path verified to exist in references/), plus the assets/cre-docs-index.md URL index; references are exactly one level deep, clearly signaled with bold 'always read before...' conditions, and the body keeps only routing and invariants inline. This matches the anchor's well-signaled, one-level-deep structure. | 5 / 5 |
Total | 17 / 20 Passed |