Content
86%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 an exemplary lean operational skill: executable commands, concrete flag semantics, explicit validation with a review checklist, and no wasted tokens. Its two real weaknesses are the absent validate-failure feedback loop and a broken progressive-disclosure chain — the findings template it defers to is not in the bundle, leaving the output format underspecified.
Suggestions
Ship templates/findings.md in the bundle (or inline the exact required section list) so the Findings format section does not depend on a file that is not present.
Add one line for the failure path, e.g., "If validate reports errors, fix the findings/sidecar and re-run validate before stopping," to close the error-recovery loop.
Verify agents/planner.md and the ../code-quality/SKILL.md link resolve from the installed skill location, or inline the one-sentence rule they carry.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense and assumes competence throughout — no padding, no explanation of what Exa or research is, and every rule carries skill-specific information the model could not infer (e.g., "Provider payloads go in the sidecar JSON (findings.raw.json beside the report by default), never inline"). This matches 'Lean and efficient; every token earns its place'; the only near-miss is the small auto-generated kendex marker comment block, which is not enough to drop to 4's 'minor instances of over-explanation'. | 5 / 5 |
Actionability | The Running and Validation sections give five copy-paste-ready commands (report with --mode/--output, --query-file, json, validate, doctor), the EXA_API_KEY bullet specifies exactly where to set it including the op:// fallback, and flag semantics are concrete ("--include-domain github.com admits every repo on it and excludes everything else"). This matches 'Fully executable; copy-paste ready code or commands'; the only unspecified invocation is the Pi web_research path, which is a harness-delegated case rather than a gap in the skill's own commands. | 5 / 5 |
Workflow Clarity | The sequence is explicit — run the report in a chosen mode, write findings to the caller's path, then "run validate and stop" — with a concrete validation checkpoint (exit 0 semantics, what validate checks) and a human-review checklist ("Claims the cited sources contradict... Off-topic sources that share an acronym"). It falls short of 5 because there is no explicit error-recovery loop: the body never says what to do when validate reports errors, unlike the anchor-5 pattern 'If errors: fix and re-validate'. | 4 / 5 |
Progressive Disclosure | The overview-to-detail split is good in principle — flag declarations and limits are correctly pushed into scripts/deep-research (verified present, including modeDefaults/parseArgs/applyMode) — but scored against the actual bundle, two referenced paths dangle: templates/findings.md ("carries exactly the sections validate requires") does not exist in the bundle, and agents/planner.md is likewise unverifiable. Since the Findings format section is thin precisely because it defers to a missing file, navigation breaks where it matters most, fitting 'references present but not clearly signaled / could be better organized' better than 4's 'minor organization gaps'. | 3 / 5 |
Total | 17 / 20 Passed |