Content
67%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-structured, expert-oriented instruction skill with genuinely executable workflow commands and present validation checkpoints for a batch operation. Its weakness is token efficiency: substantial design rationale and a report-reading section duplicate what the runbook reference carries, and a few referenced paths (evals/, harness.config.json) do not resolve within the bundle.
Suggestions
Trim the "Reading results" section to the integrity-first rule and the definition of pass_k, and delegate the per-target manual-observation enumerations (inquire/grasp/conduct lists) to references/runbook.md, which already claims to cover how to read the report.
Introduce the `report` command concretely in the Workflow block (how it is invoked and where its output lands) instead of only referencing it in the failure-mode sentence.
Move design rationale that is not needed at run time — the sham-arm literature justification and the credential-rotation/symlink explanation — into references/runbook.md or references/grader-design.md, keeping only the operational rule in the body.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is information-dense and assumes Claude's competence (no basic-concept explanations), but it carries notable duplication and over-long rationale: the multi-paragraph "Reading results" section inlines detail that the body itself says "references/runbook.md" carries, and passages like the credential-rotation and sham-arm justifications read as design documentation rather than operational instruction. Mostly efficient with sections that should be tightened or moved matches anchor 3; it avoids 2 because nothing is beginner-level padding, and misses 4 because the duplicated report-reading and rationale material is a real trimming opportunity. | 3 / 5 |
Actionability | The Workflow section gives a copy-paste bash block (cd into scripts, setup.sh/run.sh/teardown.sh with REALIZE_RUNNER variants and env vars) and names concrete config keys (harness.config.json, REALIZE_ARMS, targets registry). Minor gaps keep it from 5: the `report` command is referenced ("makes `run` or `report` exit non-zero") but never introduced, and generating/viewing the report is never concretely shown. | 4 / 5 |
Workflow Clarity | The run sequence is clearly ordered (setup → run → teardown, with runner and arm variants), and validation is present despite the batch nature of the matrix: "Read `integrity` first", non-zero exits on "a launch failure, missing cell, unreadable predicate, or treatment-integrity failure ... while preserving partial transcripts", and explicit guidance that low-integrity rows are not evidence. It is not 5 because prerequisites (read the runbook first, install the ink-figure marketplace) are scattered across prose rather than presented as an ordered checklist with error-recovery loops. | 4 / 5 |
Progressive Disclosure | Sections are well organized and the core references — references/runbook.md, references/grader-design.md, scripts/harness.mjs — exist in the bundle, are one level deep, and are clearly signaled both inline and in an annotated "Additional Resources" list. Minor gaps: the body references `evals/*.sh` and `harness.config.json`, which are not present in this bundle, and the long per-target report-reading enumerations are inlined where the runbook reference already covers them. Good structure with minor organization gaps matches anchor 4 rather than 5. | 4 / 5 |
Total | 15 / 20 Passed |