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.
A well-engineered, highly actionable workflow: deterministic-first bash with explicit abstain handling, pinned machine-checkable output formats, and a clearly sequenced six-step process with strong validation checkpoints and fallback paths. The main weakness is redundancy — the context-budget and no-content-reading constraints are stated three times across the Context Budget, step 3, and Important sections.
Suggestions
Consolidate the repeated constraints: state the context-budget rules (no README/source reading, name+description only, tightly scoped remote queries) once and reference them from steps 3 and Important instead of restating them in three places.
Provide the concrete search-filtered query for groups with more than 50 repos (e.g. a glab/gh search command filtered by ticket keywords) and an explicit URL-construction example for the clone step.
Move the two pinned output templates (Repo Match Results and Resolved Repository blocks) into a short reference file to slim the main body while keeping them machine-checkable.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body assumes Claude's competence (no explaining git, jq, or VCS concepts) and its tokens are mostly operational, but the same constraints repeat across three sections — "Never read full READMEs, source files" (Context Budget), "no cloning, no README reading, no file inspection" (step 3), "never read repo contents, READMEs, or source files" (Important) — and the sub-agent note appears at both the top and the closing section. This is more than the "minor instances that could be trimmed" of the 4 anchor, but far from the padded concept explanations of the 2 anchor. | 3 / 5 |
Actionability | The skill provides copy-paste-ready bash for repo scanning, lib sourcing with abstain handling ("val=$(detect_vcs_platform \"$REMOTE\") || val="""), scoped glab/gh queries, cloning, and setup hints, plus exact pinned output templates. It is not a 5 because the >50-repo case ("use a search query filtered by ticket keywords") and URL construction ("construct URL from platform + group + repo name") are described rather than given as commands, and bitbucket/azure remote fetch has no command. | 4 / 5 |
Workflow Clarity | Steps 1–6 are explicitly sequenced with clear skip conditions ("Only reach this step if no local repo matched confidently"; a HIGH local match "does NOT skip the Repo Match Results block") and explicit validation checkpoints and recovery loops: lib abstain → AI fallback, CLI missing → "fall back to local-only results", ambiguous → "return top 3 candidates and let the caller ask the user", and mandatory user confirmation before cloning. This matches the 5 anchor (explicit validation steps, feedback loops, checklists); cloning is not a destructive/batch operation, so the cap of 3 does not apply. | 5 / 5 |
Progressive Disclosure | No bundle files exist (references/, scripts/, assets/ are all absent), so the skill is a single self-contained file with clear section headers, and its only external references (~/.claude/lib/determinism/functions/*.sh and CONTRACT.md) are one level deep and clearly signaled. It matches the 4 anchor (good structure, minor organization gaps) rather than 5 because ~200 lines are entirely inline — the output-format templates could live in a reference file — and it exceeds the simple-skill exception threshold. | 4 / 5 |
Total | 16 / 20 Passed |