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.
A dense, operational orchestration guide: exact commands, precise output and error contracts, explicit worktree isolation rules, and a fallback ladder with feedback loops make the workflows exemplary. The weaker spots are minor — Windows-only path separators reduce copy-paste readiness, and the anti-patterns table duplicates rules already stated in the flows.
Suggestions
Use forward slashes in the invocation contract and benchmark command (and keep Windows specifics in the Windows notes section) so the commands are copy-paste ready on POSIX as well as Windows.
Deduplicate the anti-patterns table against the in-flow HARD RULES and Golden Rule — either reference the rules from the table or keep only the table — trimming roughly 15 lines of restated guidance.
Move the Benchmarking and Windows-notes detail into a reference file (e.g. references/bench.md, references/windows.md) so SKILL.md stays a lean overview with one-level-deep, well-signaled references.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and operational with no concept explanations Claude already knows ("The script auto-manages the shared server: health-checks 127.0.0.1:4096...", terse HARD RULES, table-formatted prerequisites and components). Not 5: the anti-patterns table restates five rules already given in the Golden Rule and swarm flow (e.g. piping raw NDJSON, two writers per directory, worker self-merge), and the cross-references section adds trimmable sibling-skill links. Not 3: there is no padded or tutorial-style explanation anywhere. | 4 / 5 |
Actionability | Concrete, fully specified commands: the invocation contract "node \"<skill-dir>\scripts\oc-worker.mjs\" [--agent scout|coder|tester] [--dir <path>] [--model <p/m>] [--timeout 900]..." with an exact one-line JSON output contract, "git worktree add ../<repo>-wt-N -b swarm/N", "npm i -g opencode-ai", and the bench command. Not 5: paths in the invocation contract and bench command use Windows backslashes, so they are not copy-paste ready on POSIX, and the worker task prompts / slash-command internals are delegated to assets that are not shown — minor gaps matching anchor 4 ('mostly executable with minor gaps') rather than 'copy-paste ready'. | 4 / 5 |
Workflow Clarity | The swarm flow is a clear 6-step numbered sequence with explicit validation ("Review every diff yourself... YOU are the only merger", "Merge approved branches, remove worktrees, run tests"), the /oc flow requires showing "git diff before letting the user commit", and the fallback ladder is an explicit feedback loop ("Worker ok:false -> re-invoke once... Still failing -> orchestrator performs that subtask inline... opencode entirely down -> abandon workers") with enumerated failure stages (args/exec/api/parse/empty). HARD RULES act as a checklist — matching anchor 5's explicit validation, feedback loops, and checklists. | 5 / 5 |
Progressive Disclosure | Well-organized sections plus a components table indexing every bundle path (scripts/oc-worker.mjs, oc-status.mjs, oc-aggregate.mjs, bench/run-bench.mjs, bench/grader-prompt.md, assets/commands/, assets/agents/) — all one level deep with no nested references, and slash-command/agent definitions are properly kept in assets rather than inlined. Not 5: no bundle files were provided to verify the referenced paths, and the ~165-line body carries detail (Benchmarking, Windows notes) that would fit better in a split-out reference file; not 3: what is inline is orchestration rules that must stay in context, and navigation is easy via the table. | 4 / 5 |
Total | 17 / 20 Passed |