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.
A well-structured orchestration skill body with an exceptionally clear sequenced workflow, explicit validation gates, and well-annotated one-level references. Its main weaknesses are repetition of the registry/scope-guard disclaimers across three sections and delegation of the registry-proposal mechanics to out-of-bundle references without any inline example.
Suggestions
State the registry propose rule once (e.g., in Save Results or the Skill Contract Writes bullet) and have the intro scope guard simply link to launch-registry, cutting the near-verbatim repetitions that cost conciseness.
Show one concrete minimal example of the `operation: propose` request to `registry-events.py` (or the `memory/events/launches.ndjson` payload shape) inline so step 7 is executable without chasing external references.
Collapse the triple-listing of roi-calculator / report-generator / performance-analyzer (intro scope guard, Reference Materials, Next Best Skill) into a single Scope/Reference section, and verify that `scripts/connectors/*.py` paths actually resolve from the deployed bundle.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | Mostly efficient — no explanations of concepts Claude already knows and dense, information-bearing prose — but noticeably repetitive: the registry disclaimer ("never writes memory/launch-registry/ records directly... submits the outcome snapshot... via an authorized operation: propose request to registry-events.py") appears nearly verbatim in the intro, the Writes bullet, and Save Results, and roi-calculator / report-generator / performance-analyzer / momentum-planner are each re-linked three times (intro scope guard, Reference Materials, Next Best Skill). This is more than the minor trimmable instances of the level-4 anchor, fitting "could be tightened" at level 3. | 3 / 5 |
Actionability | For an instruction-only skill the guidance is largely executable: 8 concrete numbered steps, exact save paths (`memory/launch/launch-retro-analyzer/YYYY-MM-DD-<launch-or-product>-retro.md`), explicit output schemas (truth vs reference columns, Measured/User-provided/Estimated labels, `retro_status: PROVISIONAL | NEEDS_INPUT`), and ready-to-use Quick Start prompts. Minor gaps keep it below 5: the registry proposal is never shown concretely — the actual `registry-events.py` invocation or ndjson payload shape is delegated entirely to references — so the reader cannot execute step 7 without chasing them. | 4 / 5 |
Workflow Clarity | The 8-step sequence has explicit validation checkpoints and feedback loops: step 1 gates on complete receipts ("Missing or partial receipts keep the launch join open and the retro provisional"), step 5 emits provisional recommendations when evidence is incomplete, step 7 notes registry acceptance "does not manufacture missing receipts", and step 8 blocks progression until the retro is terminal, otherwise routing the missing-receipt list back to launch-monitor. The "Done when" contract and explicit PROVISIONAL/NEEDS_INPUT statuses act as a checklist, matching the level-5 anchor; the writes are proposals gated on user confirmation, so the destructive-operation cap does not apply. | 5 / 5 |
Progressive Disclosure | Good structure: a Reference Materials section signals each external file with a one-line purpose annotation, references are one level deep, and sibling-skill hand-offs are clearly labeled. It falls short of the level-5 anchor: no bundle files exist in this skill's directory, yet the body references paths like `scripts/connectors/hn.py` and deeply relative repo paths (`../../../references/...`, `../../../CONNECTORS.md`) that live outside the bundle, so navigation depends on unverifiable external structure; the connector details in Data Sources could also live in a reference file. | 4 / 5 |
Total | 16 / 20 Passed |