Content
56%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 lean, well-sectioned process skill with a genuine validation-and-stop discipline and one concrete lint command, but its guidance relies heavily on reference material that is absent from the bundle, and it duplicates rules across Constraints, Anti-patterns, and Gotchas. Fixing the reference wiring and consolidating the redundant guardrail sections would lift it substantially.
Suggestions
Make every reference resolvable: link the actual bundle file (e.g. `references/context-preservation.md` for the context-disposition policy) from the body, and either include the referenced files (execution-modes, handoff-and-shipping) or drop/repair their paths, replacing the `repo:` prefix with a consistent relative path.
Add concrete mechanics for the abstract steps: what the three execution lanes are (or a one-line pointer to a resolvable file), and the exact commands/sequence used to resolve and update the Linear work item, mirroring the specificity of the traceability lint command.
Merge Constraints, Anti-patterns, and Gotchas into a single non-duplicating section (each rule currently appears 2-3 times) and drop the "When to use" section that restates the frontmatter description.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body avoids explaining concepts Claude already knows, but carries noticeable duplication: "Do not claim completion without validation evidence" and "Do not expand scope beyond approved artifacts/IDs" each appear in both Constraints and Gotchas, the "When to use" section restates the frontmatter description, and the Failure mode section repeats the Validation stop rules. Mostly efficient, but it could be tightened by merging Constraints/Anti-patterns/Gotchas. | 3 / 5 |
Actionability | There is one concrete executable command ("run `python3 Infrastructure/scripts/validation-and-linting/he_linear_traceability_lint.py <artifact-path>`") and specific stop conditions, but most guidance is policy-level ("Choose the execution lane", "Resolve the active Linear work item", "Implement one verified slice at a time") without the concrete steps to execute them. The mechanics are deferred to reference files that do not exist in this bundle, so the guidance is incomplete rather than executable. | 3 / 5 |
Workflow Clarity | The Procedure gives a clear 4-step sequence (choose lane, resolve Linear/IDs/gates, implement verified slices, update handoff), and the Validation section supplies explicit checkpoints including "Stop at the first failed gate" and a drift feedback loop ("Reflect contract drift in the governing artifact before off-plan implementation continues"). Minor gaps remain: steps 1-3 have no concrete commands, and slice-level verification mechanics are unstated. | 4 / 5 |
Progressive Disclosure | The body is well-sectioned and keeps itself to overview length with a clearly labeled References section, which is the right instinct. However, judged against the actual bundle structure: none of the 7 referenced paths (e.g. `Plugins/harness-engineering/fixtures/preserved-context/skills/he-work/SKILL.full.md`) exist in the bundle, one reference uses a non-standard `repo:` prefix, and the one real bundle file (`references/context-preservation.md`) is never referenced from the body, leaving it undiscoverable. Structure is present but navigation is broken. | 3 / 5 |
Total | 13 / 20 Passed |