Content
77%Reviews the quality of instructions and guidance provided to agents. Good implementation is clear, handles edge cases, and produces reliable results.
This is a strong, production-grade skill with excellent actionability and workflow clarity — it gives Claude precise tools, commands, decision criteria, and a well-bounded retry/escalation model. Its main weakness is moderate verbosity: some sections repeat constraints already embedded in the workflow, and the untrusted-content boilerplate is lengthy. The inline structure is acceptable given no bundle, but could benefit from slight reorganization.
Suggestions
Trim the 'Untrusted content' section to 2-3 sentences — Claude understands prompt injection defense; the current block is ~6 lines of elaboration on a single concept.
Consolidate the 'Do not' section by removing items already stated as constraints in the workflow steps (e.g., 'do not merge on red' is already enforced in Step 5's preconditions).
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The skill is mostly efficient and project-specific, but includes some sections that could be tightened — the 'Untrusted content' block is lengthy boilerplate, and some explanations (e.g., the tool conventions rationale citing a specific GitHub issue) add tokens without proportional value. The 'Do not' section partially repeats constraints already stated in the workflow steps. | 2 / 3 |
Actionability | Highly actionable: specifies exact MCP tool names, exact CLI commands (pnpm lint && pnpm db:check:migrations && ...), exact merge strategy (squash), exact commit format (Conventional Commits), and concrete decision criteria (3 fix attempts, all-green required checks). Every step tells Claude precisely what to do and with which tool. | 3 / 3 |
Workflow Clarity | Excellent multi-step workflow with clear sequencing (Steps 1-6), explicit validation checkpoints (run local gate before push, poll CI, verify post-merge linkage), a bounded retry loop (3 fix attempts), and a well-defined escalation/feedback path (Step 6). The validate-fix-retry loop for CI failures is explicitly modeled. | 3 / 3 |
Progressive Disclosure | The content is well-structured with clear headers and numbered steps, and references external documents (CONSTITUTION, AGENTS.md, ADRs, SPECs) appropriately. However, with no bundle files provided, all content is inline in a single file that runs fairly long (~100 lines of dense instruction). Some sections like the 'Do not' list and 'Untrusted content' block could potentially be factored out, but the skill is not egregiously monolithic for its complexity. | 2 / 3 |
Total | 10 / 12 Passed |