Content
73%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 strong orchestrator skill body: an unambiguous phased workflow with exceptional validation gates and concrete commands throughout. Its weaknesses are token efficiency — the same rules are restated across multiple sections — and a bundle that ships only SKILL.md while deferring key procedures to rule files that are not present alongside it.
Suggestions
State each --resolve-all disposition rule once (the Mode Detection table plus Hard Rules) and let Phases 3/5/6 reference it, rather than restating the full behavior in every section.
Cite 'rules/handoff.md#worker-prompt-template' once in Phase 6 and remove the two near-verbatim repeat citations in the same section.
Move the full outcome-emit memory.write schemas fully into rules/self-improvement-loop.md (or the shared rule file) and keep only a one-line pointer plus the verdict/relevance table in SKILL.md, eliminating the duplication between the 'Outcome emit' and 'Self-Improvement' sections.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is mostly dense and purposeful, but repetition is noticeable across the ~540 lines: the --resolve-all disposition rules are restated in at least four places (Mode Detection table, Phase 3 treatment table, Phase 5 pack section, Phase 6 worker pass, Hard Rules), the link 'rules/handoff.md#worker-prompt-template' is cited three times in Phase 6 alone, and the outcome-emit memory-write details appear nearly verbatim in both the 'Outcome emit' section and the 'Self-Improvement' section. Not 4: the duplication is more than minor trimming — consolidating to one canonical statement per rule (plus the Hard Rules) would cut a meaningful fraction of the document; not 2: there is no padding explaining concepts Claude already knows, and most sections earn their tokens. | 3 / 5 |
Actionability | Concrete, executable guidance dominates: exact gh calls ('gh pr view --json number,state,headRefName,headRefOid,isDraft', 'gh api repos/<owner>/<repo>/pulls/<n>/reviews'), 'gw checkout <pr-url-or-number>', 'git status --porcelain' clean-state checks, a full /confidence decision matrix, a copy-ready jq/grep one-liner for won't-fix language, and shaped memory.write calls. Not 5: several load-bearing specifics (worker prompt template, input parsing grammar, gate procedures) are deferred to rules/*.md files that are not present in this bundle, and the Agent() dispatch is illustrative rather than copy-paste ready — minor gaps remain for a reader with only SKILL.md. | 4 / 5 |
Workflow Clarity | The eight phases are explicitly sequenced with per-phase outputs, and validation checkpoints are everywhere: refuse MERGED/CLOSED PRs, worktree clean-state + HEAD==headRefOid verification, the two-gate /critical → /confidence validation with a numeric decision matrix and a non-removable Must-fix override, contradiction detection before dispatch, worker pre-push checks, non-fatal failure handling for resolve passes, CI-red as an explicit stop reason under --watch, and a Phase 7 report that audits applied/surfaced/skipped/resolved. Error-recovery and feedback loops (push retries, watch-loop baseline advancement) match the 5 anchor. | 5 / 5 |
Progressive Disclosure | The file declares itself a 'thin index' and delivers: a 'Source of truth' note, a Rules table mapping each rule file to the phase when it loads, phase-directed load instructions ('load it now when --watch is set'), and one-level-deep references to rules/*.md and templates/*.md. Not 5: some detailed content that arguably belongs in those rule files is inlined and duplicated (the outcome-emit memory schemas, --resolve-all dispositions), and several references point outside the skill directory to ../../../agents/shared/rules/*.md, while the bundle itself ships only SKILL.md — the referenced files cannot be verified as present. Not 3: the structure, navigation, and signaling are clearly good and the heavy machinery is genuinely split out. | 4 / 5 |
Total | 16 / 20 Passed |