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.
An unusually dense, expert-level conventions document: every rule is actionable, environment-specific, and backed by named regression tests, with no filler or basic-concept padding. Its main weakness is structure — the two longest topics (Windows CI sharding, the macOS Go race bug) are inlined as single monolithic bullets instead of being split into referenced files.
Suggestions
Move the Windows CI sharding details and the macOS Go 1.26 race-bug diagnostics into separate reference files (e.g. references/windows-ci.md, references/macos-race-bug.md), keeping a one-line rule plus a pointer in SKILL.md.
Trim historical narrative such as issue numbers and timeout-growth history, or consolidate it into a short 'rationale' note in the reference file.
De-duplicate the 'Regressions:' test lists at the end of long bullets where the same test names are already cited inline.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Nearly every token carries non-inferable repo-specific fact (env var names, package paths, shard topology, regression test names) with zero explanation of concepts Claude already knows. Minor trimmable material remains — issue-number history ('steps grew back to the 15m timeout in issue #1580, daemon/gate in #1682') and the trailing regression-test list that repeats names already cited inline — which keeps it at anchor 4 rather than the fully lean anchor 5. | 4 / 5 |
Actionability | Guidance is fully concrete and executable in instruction form: exact env vars (GIT_CONFIG_COUNT, GIT_CONFIG_NOSYSTEM=1, GIT_CONFIG_GLOBAL, NM_HOME, NM_E2E_DAEMON_MAX), commands (make e2e, scripts/e2e.sh), named helpers to copy from (stepstest.Init / LinkFakeCLI, internal/testgit.RealGit), and explicit prohibitions ('Do not re-exec the race-instrumented test binary as those names'). Per the rubric's instruction-skill note, the absence of code blocks is not penalized when guidance is this specific. | 5 / 5 |
Workflow Clarity | This is a conventions list rather than a sequenced workflow, and each rule states the convention, its rationale, and the enforcement test that fails on regression (e.g. shard tests 'fail when a declared test or a first-character probe name runs in zero rows or in more than one'). Rules are unambiguous with named enforcement checkpoints (anchor 4), but there is no guided sequence across the many topics, so it does not reach anchor 5. | 4 / 5 |
Progressive Disclosure | No bundle files exist and all content is inlined in one SKILL.md under only two section headers. The ~350-word Windows sharding bullet and the ~250-word macOS Go toolchain-bug diagnostic are self-contained topics that clearly belong in separate reference files, matching anchor 3 ('content that should be separate is inline'). It is not anchor 2 because the headers and per-topic bullets do provide genuine structure. | 3 / 5 |
Total | 16 / 20 Passed |