Content
63%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 highly actionable — exact commands, jq templates, decision tables, and a full output format — with a well-sequenced six-step workflow and explicit validation gates. Its weaknesses are token efficiency (the same guardrails restated three or more times across sections, ~450 lines that could be halved) and structure (a monolithic single file with no reference bundle despite content that would clearly benefit from splitting).
Suggestions
Split stable reference material — the Step 5 CI exit-code taxonomy, the Step 1.6 trim/AOT rules, and the Review Output Format template — into files under references/ (e.g. references/ci-semantics.md, references/maui-conventions.md) and keep SKILL.md as a lean workflow overview with clearly signaled links.
Deduplicate the guardrails: state each rule once (anonymous-retrieval fallback, never-LGTM-on-unverified-CI, prior-❌-override) in a single canonical section and reference it by rule number elsewhere instead of restating the full prose in Core Principles, Steps 1/4/5, and the Verdict Consistency Rules.
Tighten long prose paragraphs (e.g. the Step 6 blast-radius and non-inert-prose paragraphs, and the exit-code case (b) spelling of both stderr variants) into compact table rows or bullet conditions to cut tokens without losing the decision logic.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The ~450-line body restates the same guardrails repeatedly: anonymous-retrieval fallback appears in Core Principle 7, Step 1, and Step 4; 'never LGTM on red/pending CI' appears in Step 5, the confidence-cap table, and Verdict Consistency Rule 6; the CI exit-code taxonomy and blast-radius prose are heavily padded. This matches 'noticeably verbose; several unnecessary explanations or padded sections' — it is clearly not the mostly-efficient anchor 3, though it avoids teaching concepts Claude already knows. | 2 / 5 |
Actionability | Quotes: 'curl -fsSL -H 'Accept: application/vnd.github.patch' https://api.github.com/repos/dotnet/maui/pulls/<PR_NUMBER>', the full 'gh api repos/dotnet/maui/pulls/$PR_NUMBER/reviews --method POST ... --input <(jq -n ...)' posting command, the three gh/jq reconciliation commands, and the complete copy-paste output template. Commands are copy-paste ready and cover the common live-PR and materialized-input cases, matching the fully-executable anchor. | 5 / 5 |
Workflow Clarity | Quotes: 'Step 1: Gather Code Context (No PR Narrative)' through 'Step 6: Blast Radius, Failure-Mode Probing, and Verdict', with validation checkpoints ('Verify whether each ❌ Error finding was addressed', 'Classify based on the stdout row content ... not the exit code alone') and a Completion Criteria checklist. Clear sequenced workflow with explicit verification, but interleaved fallback prose inside Step 1, duplicated rules across sections, and branching (materialized vs live input) make the sequence harder to follow than the top anchor. | 4 / 5 |
Progressive Disclosure | The bundle contains no references/, scripts/, or assets/ directories — everything is inlined in SKILL.md. Section headers are clear and the body points to repo-external files ('.github/agents/maui-expert-reviewer.md'), but ~150 lines of CI exit-code semantics, trim/AOT evidence rules, and the output-format template clearly belong in separate reference files. This matches 'some structure but ... content that should be separate is inline'; it is not anchor 2 because headers and navigation are present and consistent. | 3 / 5 |
Total | 14 / 20 Passed |