Content
75%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 highly actionable, well-structured skill body with real commands, code, and file paths, plus a properly signaled one-level-deep reference for host-bridge internals. Its main weakness is length and repetition — the same directives (connect command, link purity, embedApp usage) appear in Rule, How, Do, and Don't sections, inflating token cost without adding information.
Suggestions
Deduplicate directives that appear in both the How sections and the Do/Don't lists (e.g. the `npx ... connect` command, `link` purity, `embedApp()` usage) — state each once and cross-reference it.
Move the detailed catalog-policy prose in §1a (compact catalog, authenticatedReads, db-schema/db-query exposure) and the provider API connector section into a reference file, keeping only the decision rules inline.
Tighten §1b fast-path expectations into a compact bullet table of intent → expected action, removing the surrounding prose about host plumbing.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The content is dense with framework-specific knowledge (no padding with concepts Claude already knows), but directives repeat across sections: the `npx @agent-native/core@latest connect` command appears in Rule, How §1, and Do; `link` purity rules in §2, Do, and Don't; `embedApp()` guidance in §2a, Do, and Don't. Several long prose blocks (e.g. §1a catalog policy, provider connectors) could be tightened or moved to references. Not score 2: there are no explanations of concepts Claude already knows and every section carries real, non-generic information. | 3 / 5 |
Actionability | Fully executable guidance throughout: copy-paste-ready commands (`npx @agent-native/core@latest connect <url>`, `reconnect`, `pnpm exec agent-native mcp install --client claude-code`), concrete URLs, and real code examples with file paths (mail's `manage-draft.ts` `link` builder, `embedApp()`, `callMcpTool`) covering the common cases. | 5 / 5 |
Workflow Clarity | The How sections give a clear numbered sequence (1 connect → 2 link builder → 2a mcpApp → 3 open route → 4 ingest actions → 5 local dev) with fast-path expectations (§1b) and useful diagnostic guidance ("A sparse or empty `tools/list` is diagnostic, not proof of auth failure"; "validate with a fresh tool call" after descriptor changes). Not score 5: validation checkpoints are scattered across prose and Do/Don't lists rather than explicit per-step checkpoints with fix-and-retry loops. | 4 / 5 |
Progressive Disclosure | There is exactly one one-level-deep reference (`references/mcp-apps-embedding.md`, verified to exist) and it is clearly signaled with a purpose ("Read it before changing the `embedApp()` shell, the `ui://` resource, or any host-bridge behavior"). Not score 5: the ~550-line inline body inlines several subsystems (fast-path expectations, app-facing MCP connectors, provider API connectors, blueprint installer) that could plausibly be split into references, and several other docs are named only implicitly. | 4 / 5 |
Total | 16 / 20 Passed |