Content
81%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 disciplined instruction skill: every phase has explicit gates, the guidance is copy-ready concrete, and detail is genuinely delegated to eight real reference files with imperative read-now signals. The two costs are embedded justificatory prose that inflates token load and Phase 4's routing detail, which sits inline where the other phases push detail to references.
Suggestions
Trim rationale sentences that explain why a rule exists (e.g. "Every option loses something the agent cannot choose on the user's behalf, which is why this question survives"; the "Two facts make it less obvious than it looks" framing) — the rules are enforceable without the justification, saving meaningful tokens in the always-loaded body.
Move the question-2 ships/stays-local decision tree and its PR-capability facts into references/post-fix-handoff.md (or a new routing reference), keeping only the two questions and the three outcomes in the body — this matches how Phases 0-2 and 3 delegate detail.
Collapse the Artifact Root HTML comment markers and validation bullets into the two or three operative lines, or relocate the validation detail to a reference, since it is rarely needed per run.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body assumes Claude's competence throughout (no git/git-hub primers, dense imperative prose) but embeds substantial rationale that could be trimmed, e.g. "Every option loses something the agent cannot choose on the user's behalf, which is why this question survives" and the extended already-pushed-vs-offered explanation in Phase 4's question 2. Not 4 because these over-explanations are more than minor; not 2 because there is no concept-teaching padding and every section drives behavior. | 3 / 5 |
Actionability | Fully concrete instruction-only guidance: exact commands with pitfalls handled ("`git rev-parse --abbrev-ref origin/HEAD` with its `origin/` prefix stripped"), a copy-ready Debug Summary template, enumerated fix-choice options, exact status spellings ("fixed-and-pushed | fixed-not-pushed | diagnosed-no-fix | flaky-infra | needs-human"), and precise delegation rules to named skills. Per the rubric's instruction-only note, the absence of code is not penalized when guidance is this actionable. | 5 / 5 |
Workflow Clarity | Five phases are sequenced in order with explicit validation gates: the causal-chain gate before Phase 3 ("'Somehow X leads to Y' is a gap"), the same-turn presentation-before-gate rule, escalation triggers ("2-3 hypotheses exhausted... or 3 failed fix attempts"), and pre-commit confirmation of unstaged user work. Feedback loops and recovery rules are stated or delegated to named references; nothing in the risky commit/push path lacks a checkpoint. | 5 / 5 |
Progressive Disclosure | All referenced files exist in references/ and are one level deep, clearly signaled with directives ("Read references/investigate.md now and follow it for Phases 0-2"), and the body deliberately keeps only the gates inline. Not 5 because the body is more than an overview: Phase 4's ~20 lines of routing decision logic (ships/stays-local/not-a-git-repo, PR-capability facts) reads as reference-grade detail inlined in SKILL.md rather than split out. | 4 / 5 |
Total | 17 / 20 Passed |