Content
77%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.
The body delivers exceptional operational specificity — executable commands, explicit validation checkpoints, error-recovery guidance, and well-sequenced workflows for both lanes. Its weaknesses are redundancy (the deadman doctrine repeated across four sections inflates the ~500-line body) and the absence of any progressive-disclosure structure, with separable appendix-level content inlined in one monolithic file.
Suggestions
State the deadman/durable-execution rules once in a canonical section and reference it from the Contract and Anti-Patterns lists instead of restating them in four places — this alone would cut a substantial fraction of the body.
Move the stage-checkpoint appendix and the detailed shell-job preconditions into separate reference files (e.g., references/stage-checkpoints.md, references/shell-jobs.md) with one-line pointers from SKILL.md, giving the skill a one-level-deep reference structure.
Consolidate the scattered repo-file mentions (src/core/minions/handlers/*.ts, docs/guides/minions-shell-jobs.md, skills/conventions/*.md) into a clearly signaled references section so navigation targets are discoverable in one place.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with project-specific operational detail and explains nothing Claude already knows, but the deadman/durable-execution rules are restated four times (Contract, Durable execution intro, "Failure modes the pattern must own", Anti-Patterns) and Rung 1 re-restates the shell-job Preconditions. Mostly efficient with clear tightening opportunities, matching anchor 3 rather than anchor 4's "minor instances". | 3 / 5 |
Actionability | Guidance is copy-paste ready throughout: exact CLI invocations ("gbrain jobs submit shell --params '{\"cmd\":\"echo hello\",\"cwd\":\"/abs/path\"}'"), MCP tool calls with parameters, flag lists with sources, a timeout table, and specific error-code remedies. The only flexibility ("whatever one-shot scheduler the host offers") is explicitly justified, so the level-5 anchor holds. | 5 / 5 |
Workflow Clarity | Multi-step flows are clearly sequenced (Phases 1-5 for subagents; Preconditions → Submit → Monitor → Control for shell jobs; Rungs 1-3 for the capability ladder) with explicit validation ("run gbrain jobs stats to confirm the worker is registered", checkpoint freshness checks, the deadman reported-state check), feedback loops in "When it fails" (rate_limited → back off; coalesced → monitor, don't resubmit), and checklist-style eval contracts. This matches the level-5 anchor including its batch/destructive validation requirement (consent before paid work, stats check before >5 concurrent agents). | 5 / 5 |
Progressive Disclosure | No bundle files exist (no references/, scripts/, or assets/), and the ~500-line body is a single monolith where the content-addressed stage-checkpoint appendix, the shell-job preconditions, and the detailed doctrine clearly belong in separate reference files. Section headers and tables provide real structure (ruling out anchor 2's wall of text), but repo-file references are scattered inline rather than clearly signaled, matching anchor 3. | 3 / 5 |
Total | 16 / 20 Passed |