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.
An unusually concrete, well-gated operational procedure: executable commands throughout, explicit validation and fail-closed checks, and clear iteration/stop semantics. It is dragged down by baroque run-on prose with repeated protocol text, a non-executable spec for the safety-critical subagent spawn, and a progressive-disclosure structure whose referenced sidecar files and scripts are entirely absent from the bundle.
Suggestions
Ship the referenced bundle files (LANE.md, RISK.md, RELEASE-BOUND.md, REPEATED-FAILURE.md, DISMISSALS.md, NO-LENS.md, CHECK-TABLE.md, and scripts/trace.sh, scripts/vocab.sh, scripts/agents.lib.sh, scripts/vocab.config.sh) or remove the references — every sidecar link currently dangles.
Move the ~80-line fence shell block (fetch_bodies, vocab_checker, span_ok, typed_return_ok, checked_returns) into a scripts/ file and reference it, keeping only the call sequence and the shape contract in the body.
Rewrite the multi-hundred-word run-on paragraphs (subagent spawn spec, stop-condition bullet 1, the 'check fails closed' prose) as numbered sub-steps or bullet lists, and state the Trace-Run/Trace-Spawn hand-off protocol once instead of twice.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | No filler explaining things Claude already knows — nearly every sentence carries repo-specific policy — but several passages are conspicuously over-written: the ~600-word single-paragraph subagent-spawn spec ("A tool-restricted subagent reads those files..."), the stop-conditions bullet 1 run-on, and the fence prose at "The check fails closed". The Trace-Run hand-off protocol is stated twice nearly verbatim (step 1 and step 2), and the vocab_checker discovery rule is explained both in code comments and again in prose. Not 2 because there is no generic padding; the density is high but the tightening opportunities are real. | 3 / 5 |
Actionability | Mostly copy-paste executable: prereq guard script, gh api snapshot commands with jq filters, a GraphQL review-thread query, five complete shell functions (fetch_bodies, vocab_checker, span_ok, typed_return_ok, checked_returns), commit templates, reply endpoints, and an output-format template. Not 5: the subagent reader spawn — the core of the trust fence — is described only in prose with no executable command ("How an agent harness withholds those tools is the adapter's, not this skill's, to say"), leaving the most safety-critical step non-concrete. | 4 / 5 |
Workflow Clarity | Clear numbered sequence (prereqs → snapshot → review → triage → act → wait → stop) with genuine validation checkpoints: worktree/branch/staleness aborts (exit 2), the fail-closed checked_returns gate, explicit convergence stop conditions, and a 5-iteration stuck cap — well above the batch/destructive cap concern. Not 5: step 5's structure is muddled ("Two modes:" followed by "Either way..." before the modes are listed) and stop-condition bullet 1 buries the re-review/verdict protocol in a several-hundred-word run-on, degrading followability at exactly the decision points that matter most. | 4 / 5 |
Progressive Disclosure | The body is structured as an overview with clearly signaled one-level sidecar references ([RELEASE-BOUND.md], [DISMISSALS.md], [CHECK-TABLE.md], [LANE.md], [RISK.md], [NO-LENS.md], [REPEATED-FAILURE.md], scripts/trace.sh, scripts/vocab.sh) — but the bundle ships none of them: no references/, scripts/, or assets/ directories exist, so every sidecar link dangles. Additionally ~80 lines of fence shell are inlined in the body where the scripts/ file it names would go. Not 4: references cannot actually be navigated and content that belongs in a separate file is inline. | 3 / 5 |
Total | 14 / 20 Passed |