Content
81%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 with an explicit workflow and a strong self-check checklist, and it makes good use of two real one-level-deep reference files. Its main weakness is repetition: the courier rule and hook-blocked behavior are each explained in multiple sections, inflating token cost without adding guidance.
Suggestions
State the 'courier, not co-author' verbatim-relay rule once (in the Output contract section) and cross-reference it from the intro and Self-check instead of repeating it in full three times.
Merge 'Hook-blocked tool calls count as tool output' and 'Where the tool's stdout actually lives' into one section covering the blocked case, keeping a single Right/Wrong example pair rather than two parallel walkthroughs.
Move the dated benchmark note ('2026-09-06: ai-ide-runner-single-cursor-read-only …') into a reference or past-failures note; time-sensitive entries in the main body add tokens and will go stale.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Most sections are dense and substantive, but the core rule 'You are a courier, not a co-author' is stated three times (intro, Output contract, Self-check) and the hook-blocked/MOCK-prefix behavior is explained in three separate places ('Hook-blocked tool calls count as tool output', 'Where the tool's stdout actually lives', and duplicated Right/Wrong example blocks). A dated benchmark log entry ('(2026-09-06: ai-ide-runner-single-cursor-read-only …)') adds time-sensitive detail outside any 'old patterns' section. Not 2 because there is little concept-teaching fluff; not 4 because the repetition is substantial and clearly trimmable. | 3 / 5 |
Actionability | Fully executable throughout: exact binaries and flags ('claude -p "<prompt>"', 'opencode run', 'cursor-agent -p --trust', 'codex exec'), a copy-paste parallel bash script with 'wait' and PID tracking, discovery commands per runtime, and a concrete env override ('CLAUDECODE="" claude -p …'). Covers the common single-run, fan-out, and comparison cases. | 5 / 5 |
Workflow Clarity | The four-step workflow (Parse intent → Pick model → Build & run → Present) is clearly sequenced, and the 'Self-check before sending the final message' section is an explicit validation checklist with corrective feedback loops ('If no → go back and add it', 'Revert to pure relay'). Error-recovery paths for auth failures, missing binaries, and native-provider errors are specified (stop, report verbatim, ask user). | 5 / 5 |
Progressive Disclosure | Good structure with well-signaled one-level-deep references: full flag sets deferred to references/runtimes.md and the model catalogue to references/models.md (both files exist and match their descriptions). Not 5 because the 299-line main file carries extended Right/Wrong example walkthroughs and the multi-section hook-blocked narrative that would fit better in a reference or a single consolidated section — minor organization gaps. | 4 / 5 |
Total | 17 / 20 Passed |