Content
85%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, actionable orchestration skill with genuine feedback loops (verify-then-fix, snapshot/restore, base-rerun attribution, round 2) and near-zero token waste. It is let down by an incomplete bundle: the principle files steps 3–6 depend on are not shipped, which caps both actionability and progressive disclosure.
Suggestions
Ship the nine principle files the step 3 table references (test/acceptance.md, test/coverage.md, test/blast-radius.md, test/ui-ux.md, review/security.md, review/maintainable.md, review/readability.md, review/conventions.md, review/data-patterns.md) — without them the agent prompt in step 3 and the re-checks in step 6 are unexecutable.
State a fallback for a missing principle file (analogous to "No map means the principles run as written" for stack maps), e.g. check the principle's named concern yourself, so a partial bundle degrades gracefully.
Add a one-line bundle map near the table (or note which paths are required vs. optional) so a loading session can tell whether a referenced file failed to copy or was never meant to exist.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Every section is instruction, not explanation: no concept Claude already knows is taught, and rule content is delivered in minimal tokens ("Lines in new_lines.txt are new code; every other line is existing code, even in files the change edited"). The only authoring-side aside ("make-it-fast refers to these steps by number; renumber both") is a single clause; not 4 because there is no over-explanation left to trim. | 5 / 5 |
Actionability | Largely copy-paste ready: the git fetch / merge-base / mktemp / worktree block, the verbatim subagent prompt with per-group additions, " "$SKILL_DIR/scripts/new_lines.sh" "$BASE" ", and the snapshot/restore commands are all executable as written. Not 5 because steps 3–6 hinge on reading nine principle files ("test/acceptance.md" through "review/data-patterns.md") that are not present in the bundle, so the core checking guidance cannot actually be executed as shipped; not 3 because everything written in SKILL.md itself is concrete and executable, not pseudocode. | 4 / 5 |
Workflow Clarity | Seven numbered steps with explicit validation and error recovery throughout: failing specs are rerun at the base in a worktree to attribute blame, a verifier subagent scores and marks each finding new/existing, snapshots gate agent damage and fixes ("git restore --source="$PRE""), a fix that breaks something twice is reverted and reported, and round 2 re-checks only fix hunks with an explicit stop condition ("even if findings remain … list them as round-2 limit"). Matches the explicit-checkpoints/feedback-loops anchor; not 4 because the checkpoints are explicit, not implicit. | 5 / 5 |
Progressive Disclosure | The design is a lean overview with well-signaled, one-level-deep references (the principle table, "references/verifier.md", "references/report.md", "references/redesign.md", "references/stacks/rails.md", and both scripts/ files resolve), but the nine principle files the entire workflow pivots on — test/acceptance.md, test/coverage.md, test/blast-radius.md, test/ui-ux.md, review/security.md, review/maintainable.md, review/readability.md, review/conventions.md, review/data-patterns.md — are absent from the bundle, breaking navigation exactly where the detail lives. Not 4: missing referenced files is more than a minor organization gap; not 2: the files that do exist are clearly signaled and correctly placed, with nothing inlined or buried. | 3 / 5 |
Total | 17 / 20 Passed |