Content
73%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.
This is a strong, highly actionable skill body: an explicit autonomous loop, a validated state machine, per-broker playbooks, and real verification steps for destructive operations, with bundle references that are genuine and one level deep. Its weaknesses are redundancy — several policy rules (PeopleConnect exception, CAPTCHA policy, no-questions rule) are stated two to four times — and one broken verification instruction referencing a test runner and test file that do not exist in the bundle.
Suggestions
Fix the Verification section: scripts/run_tests.sh and tests/skills/test_unbroker_skill.py do not exist in the bundle — either include them or point to a command that actually runs.
State the PeopleConnect delete-wipes-suppression rule once in a single authoritative place (e.g. its site playbook or methods.md) and reference it from the other sections; the same consolidation applies to the CAPTCHA policy, which is currently spread across four sections.
Move the full blind-opt-out doctrine and per-method opt-out detail from the Batch/Procedure sections into references/methods.md, keeping only the decision rule and a pointer inline.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is dense with genuinely skill-specific operational knowledge (autonomy contract, tier system, cluster-parent ordering, CAPTCHA policy) and wastes little on concepts Claude already knows, but the same rules are restated multiple times: the PeopleConnect delete-wipes-suppression exception appears three times (Batch operation, Procedure step 5, plus its own bolded bullet), the CAPTCHA policy is covered in Prerequisites, Batch operation, Procedure step 5, and Pitfalls, and the "don't ask questions" rule is repeated in the Autonomy contract and Pitfalls. This matches the anchor "mostly efficient but includes some unnecessary explanation or could be tightened"; it is clearly not the lean score-5 anchor, nor padded enough to fall to 2. | 3 / 5 |
Actionability | The body provides exact commands with flags, a ready-to-run loop pseudocode that maps to real CLI commands, concrete env vars, a copy-paste dry-run verification snippet, and per-broker playbook references — overwhelmingly executable guidance. The gap: the Verification section cites "scripts/run_tests.sh tests/skills/test_unbroker_skill.py", but neither scripts/run_tests.sh nor any tests/ directory exists in the bundle, so the primary test instruction cannot be executed as written. This fits "mostly executable guidance; concrete code or commands with minor gaps" rather than the fully copy-paste-ready score-5 anchor. | 4 / 5 |
Workflow Clarity | The Procedure is an explicit 8-step sequence with a concrete while-loop over `$PDD next`, and the destructive/batch operations all carry validation checkpoints and feedback loops: `confirmed_removed` only after a verifying re-scan, subagent `found` claims re-verified by the parent, re-scan of cluster children after each parent confirms, `blocked` sites requeued rather than dropped, lock-timeout recovery guidance, and error paths for broker flows breaking. This matches the score-5 anchor (clear sequence, explicit validation, feedback loops for error recovery) — the destructive-operation cap of 3 does not apply because validation is thoroughly present. | 5 / 5 |
Progressive Disclosure | SKILL.md functions as an operational overview with clearly signaled, one-level-deep references that all exist (references/methods.md, references/site-playbooks.md, references/brokers/*.json including intelius.json, references/legal/drop.md), plus real scripts. It is not the score-5 anchor because the body still inlines substantial policy detail that belongs in methods.md (e.g. the full blind-opt-out doctrine, the PeopleConnect exception stated three times, and the 23-row quick-reference table), which could be split out. It is well above the score-3 anchor, whose hallmark is buried or unsignaled references. | 4 / 5 |
Total | 16 / 20 Passed |