Conciseness

Most content is dense, framework-specific guidance Claude could not know, but there is a noticeable amount that could be trimmed: internal implementation detail ('a logged-in browser session mints an `A2A_SECRET`-signed JWT carrying the caller's `sub` + `org_domain` and a unique `jti`'), time-sensitive spec prose ('MCP `2026-07-28`', '2025-era clients', RFC 9207, Client ID Metadata Documents) not placed in a deprecated/old-patterns section, and repetition (the `link` purity contract appears in Rule, §2, Do, and Don't; the `npx … connect` command appears three times). This sits between anchor 2's 'several unnecessary explanations or padded sections' and anchor 4's 'minor instances of over-explanation' — mostly efficient but needs tightening.

Actionability

Guidance is fully executable across the common cases: copy-paste commands ('npx @agent-native/core@latest connect https://dispatch.agent-native.com', 'pnpm exec agent-native mcp install --client claude-code|…|github-copilot'), real code for the `link` builder, `mcpApp`/`embedApp()`, and `callMcpTool` (grounded in actual files like 'templates/mail/actions/manage-draft.ts'), and exact config flags ('MCPConfig.builtinCrossAppTools: false', 'externalAgents.authenticatedReads: "auto"', 'mcpApp.compactCatalog: true'). This matches anchor 5; the only elisions ('// …schema, run…') are deliberate and standard.

Workflow Clarity

The How section is clearly sequenced (1 connect → 1a verbs → 1b fast-path → 2 link builder → 2a MCP Apps → 3 open route → 4 ingest → 5 local dev), with a numbered 4-step sequence for the open route and diagnostic/recovery guidance ('A sparse or empty `tools/list` is diagnostic…', 'reconnect the Claude connector or rescan/review the ChatGPT connector', 'validate with a fresh tool call'). Anchor 5 would require validation checkpoints embedded in each workflow with explicit feedback loops, whereas here they live mostly in the Do/Don't lists, so anchor 4 ('clear sequence with most checkpoints present; minor validation gaps') is the best fit.

Progressive Disclosure

The bundle has exactly one reference (references/mcp-apps-embedding.md), and it is clearly signaled with a read-trigger ('Read it before changing the `embedApp()` shell …'), one level deep. However, the ~540-line body inlines substantial material that belongs in reference files — MCP protocol-version/OAuth validation internals, provider API connector policy, per-client config file paths, catalog/exposure policy — matching anchor 3's 'content that should be separate is inline', and 'See the Blueprint Installer doc' has no path or link. It is better organized than anchor 3's example but clearly short of anchor 4's 'most content is appropriately placed'.

Specificity

The description lists multiple concrete actions — 'Connect external agents and MCP hosts … over MCP', 'round-trip artifacts back into the UI with MCP Apps and deep links', 'adding an action's `link` builder or `mcpApp`', 'wiring the `/_agent-native/open` route', 'exposing an "ingest" action', 'scaffolding apps from an external agent' — comprehensively covering the skill's capabilities. Anchor 5 ('lists multiple specific concrete actions; comprehensive coverage') is the closest match; anchor 4 would require minor coverage gaps, and none are evident.

Completeness

It explicitly answers both questions: the 'what' ('Connect external agents and MCP hosts … and round-trip artifacts back into the UI with MCP Apps and deep links') and an explicit 'Use when …' clause with four concrete trigger conditions. This matches anchor 5 exactly ('clearly and explicitly answers both what AND when with concrete trigger phrases') and exceeds anchor 4, whose 'when' is present but less specific.

Trigger Term Quality

Natural trigger coverage is comprehensive for the domain: ten host names users would actually say ('Claude, Claude Desktop, Claude Code, ChatGPT custom MCP apps, Codex, Cursor, Claude Cowork, VS Code GitHub Copilot, Goose, Postman, MCPJam') plus synonyms like 'external agents', 'MCP hosts', 'deep links', 'ingest', and 'scaffolding'. It matches anchor 5's comprehensive synonym coverage; anchor 4 would mean a few natural terms are missing, and none of the common entry phrases for this task are absent.

Distinctiveness Conflict Risk

The niche is clear (bridging external MCP hosts to the agent-native app) and the triggers — `link` builder, `mcpApp`, `/_agent-native/open`, ingest actions, external scaffolding — are specific to this skill, matching anchor 5 ('clear niche with distinct triggers; minimal conflict risk'). Adjacent skills like 'actions' or 'a2a-protocol' share some vocabulary ('MCP', 'action'), but the trigger phrases here point uniquely at the external-agent bridge, so anchor 4's 'minor overlap risk' understates the fit.