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.
This is a highly actionable, well-structured integration skill: real code, exact commands, clear sequencing, and a genuinely useful offloaded reference for MCP Apps embedding internals. Its main cost is size — a ~28KB body with duplicated passages between section 2a and the reference file, Do/Don't restatements of in-body rules, and several long detail sections that could be split into reference files to keep the SKILL.md overview lean.
Suggestions
Deduplicate section 2a against references/mcp-apps-embedding.md: keep only the mental model and embedApp() snippet in the body and move the duplicated prose fully into the reference.
Trim the Do/Don't lists to rules not already stated in How sections (e.g. the link-purity rules appear three times), or fold the body statements into the Do/Don't list only.
Move inline-heavy detail into reference files — the per-client config-file enumeration (section 5) and the provider API connector catalog are self-contained topics a reader only needs occasionally.
| Dimension | Reasoning | Score |
|---|---|---|
Conciseness | The body is overwhelmingly dense, project-specific fact (host setup steps, tool verbs, security gates) that Claude cannot know, so it avoids explaining known concepts — but at ~535 lines it carries real padding: section 2a's opening paragraphs are duplicated nearly verbatim in `references/mcp-apps-embedding.md`, the Do/Don't lists restate rules already given ("keep `link` pure and synchronous" appears in section 2 and again in Do and Don't), and version-dated details ("MCP `2026-07-28`", "2025-era clients") are inline rather than in a deprecated/old-patterns section. This matches anchor 3 (mostly efficient but could be tightened) better than anchor 4, and is far from the severe padding of anchor 2. | 3 / 5 |
Actionability | Fully executable guidance throughout: copy-paste-ready `defineAction` examples with real imports (`buildDeepLink`, `embedApp`, `callMcpTool`), exact commands (`npx @agent-native/core@latest connect https://dispatch.agent-native.com`, `mcp install --client claude-code|...`), exact URLs, and concrete config flags (`externalAgents.authenticatedReads: "auto"`, `MCPConfig.builtinCrossAppTools: false`). Anchor 5 (copy-paste ready, common cases covered) fits; the only elisions (`// ...schema, run...`) are reasonable omissions of irrelevant parts, not pseudocode, so anchor 4's 'minor gaps' understates it. | 5 / 5 |
Workflow Clarity | The How section is clearly sequenced (connect hosts → add `link` builder → add `mcpApp` → open route → ingest actions → local dev) and includes genuine validation/diagnostic loops ("validate with a fresh tool call; if the host still behaves like the old descriptor, reconnect...", "A sparse or empty `tools/list` is diagnostic... check OAuth scopes, compact-catalog filtering, and the client/server auth status"). These checkpoints live in prose rather than being wired into each step, which is anchor 4 (clear sequence, most checkpoints present) rather than anchor 5's integrated validate-and-recover structure; the skill involves no destructive/batch operations that would cap it at 3. | 4 / 5 |
Progressive Disclosure | Good structure with well-signaled, one-level-deep references: section 2a explicitly directs "documented in **`references/mcp-apps-embedding.md`**. Read it before changing the `embedApp()` shell", and that file exists and stays one level deep. However, the 535-line body keeps a lot of separable detail inline — the per-client config-file enumeration in section 5, the provider OAuth catalog, and the app-facing connector internals — where the anchor-5 pattern would offload bulk detail to references and keep a leaner overview. This lands above anchor 3 (whose references are poorly signaled) but below anchor 5's clean split, i.e. anchor 4. | 4 / 5 |
Total | 16 / 20 Passed |