Content
68%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 is a maximally lean delegation stub that is token-efficient and cleanly points one level deep to the shared skill, but it delegates everything: no task steps, no validation checkpoints for the batch cell-filling work, and no overview of what the shared skill contains. It works only insofar as the external shared skill is guaranteed to be present and loadable.
Suggestions
Add a 3-5 line overview of the shared skill's workflow (research steps, cell-filling conventions, citation format) so the stub orients the reader before delegation.
Mention where validation lives (e.g., verify cited evidence and re-run harness-probe checks after filling cells) so the batch editing work has a checkpoint even at the stub level.
Reference the shared skill via a path that is stable across environments (or bundle a copy/symlink under the skill directory), since `~/.agents/skills/...` is machine-specific and unverifiable from the bundle.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is two lines: a generated-file marker ("Generated by scripts/agent-harnesses.py; do not edit directly.") and a single delegation instruction. There is no padding and no explanation of concepts Claude already knows; every token is operational. This matches the lean-and-efficient anchor exactly. | 5 / 5 |
Actionability | "Load and follow the shared skill at `~/.agents/skills/harness-research/SKILL.md`" is a concrete, executable directive with an exact path, but it is the only guidance present — how to research the release, fill cells, or format citations is entirely absent from the body. Not 4 because the gap is the whole task detail, not a minor one; not 2 because the written instruction is a precise, executable step rather than a high-level hint. | 3 / 5 |
Workflow Clarity | The single delegation step is unambiguous, but the multi-step research-and-edit workflow promised by the description has no sequence or checkpoints in the body, and the batch operation of filling capability cells has no validation guidance, which caps this at 3. Not 2 because the one step that is written is coherent and precise. | 3 / 5 |
Progressive Disclosure | The body is a single, clearly signaled, one-level-deep pointer with trivial navigation and no nested references; no bundle directories (references/, scripts/, assets/) exist, and the pointer targets an absolute home-directory path outside the skill bundle. Not 5 because the body provides no overview or orientation — the reader learns nothing about the skill's scope before following the link. | 4 / 5 |
Total | 15 / 20 Passed |