Content
50%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.
This is a high-signal internal invariants map: every rule names its owning file/function and guarding regression tests, and the fail-closed validation posture is consistent and explicit. Its weaknesses are form, not substance — extremely dense run-on prose with inlined test-name lists, no followable workflow, and all detail living inline in SKILL.md instead of being split into per-topic reference files.
Suggestions
Split each bolded section (branchsync, push binding, force-push safety, rebase.strategy, private mirrors) into its own reference file, leaving SKILL.md as a one-screen index of invariants plus pointers — the body is ~22KB of dense material that belongs in progressively disclosed files.
Replace the long inline 'Regressions:' test-name lists with a single pointer to the owning test file (e.g. 'see internal/branchsync/recover_test.go') — the names are re-discoverable in the file and cost hundreds of tokens inline.
Break the multi-clause run-on bullets (some exceed 150 words) into short imperative rules, and add a short ordered 'when touching these files, check these invariants' checklist so the skill has a followable workflow shape.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is a ~22KB wall of dense run-on bullets, with entire regression-test name lists inlined inline (e.g. 'Regressions: TestRefreshSlowSuccessfulLsRemoteDoesNotStealFetchBudget, TestRefreshSlowButSuccessfulLsRemoteAloneExceedsItsOwnBudgetReportsOffline, ...') that belong in the test files they name. It does not explain concepts Claude already knows (all facts are project-internal), so it is noticeably padded rather than anchor-1 'severely verbose with known-concept explanations'. | 2 / 5 |
Actionability | Concrete, executable guidance throughout: exact git operations ('an atomic update-ref CAS plus read-tree -m -u, never check-then-act followed by reset --hard'), exact owners ('recoverAdoptPreserved owns the reasoning', 'Public guidance is owned by internal/skill/skill.go'), named config knobs with defaults ('Service.RemoteTimeout... default config.DefaultBranchSyncRemoteTimeout, 60s'), and explicit prohibitions ('do not unify the two prompts'). As an instruction-only skill the absence of code blocks is not penalized; minor gaps (no worked example of the relation matrix) keep it below 5. | 4 / 5 |
Workflow Clarity | The skill is an invariants/reference document rather than a sequenced process, and no followable multi-step workflow is presented — sections describe rules and evidence shapes, not an ordered procedure. Validation/fail-closed behavior is stated emphatically ('rechecks... immediately before apply', 'fail closed with manual reconciliation'), so checkpoints exist conceptually, but the sequence itself is implicit, matching anchor 3 rather than 4. | 3 / 5 |
Progressive Disclosure | There are clear bolded section headers and one-level-deep pointers into the repo ('The full relation matrix and fail-safe rules live in the Recover doc comment in internal/branchsync/sync.go', 'owned by docs/src/content/docs/concepts/gate-model.md'), and no bundle files exist to reference. However, the sheer bulk of dense per-subsystem material (~22KB inline) clearly belongs in split reference files, which caps it at anchor 3 ('content that should be separate is inline') rather than 4. | 3 / 5 |
Total | 12 / 20 Passed |