Three tools for building AI-assisted mobile apps: skills, MCP servers, and the AI workflow
Key Points
- Use SKILL.md files for instructions
- Connect an MCP only when you need external data/actions
- Keep conversations under ~50% context and branch threads
Summary
This article explains the three practical tools you need to build mobile apps with AI: skills (SKILL.md instruction files), MCP (Model Context Protocol) servers to give the agent access to external systems, and an AI workflow that keeps conversations focused and context-efficient. Recommended order: add skills first, connect one MCP when you need external data or actions, then actively manage conversation context to avoid bloat.
Key Points
-
Skills
- A skill is a folder containing a single required SKILL.md; the agent reads it only when the task calls for it (progressive disclosure).
- Install via the Expo skills registry (e.g.,
npx skills add -s "*") and scope to the project unless you need global reuse. - Invoke explicitly with a slash command (e.g.,
/expo-ui) or let the agent autonomously pick relevant skills via their short descriptions. Avoid hoarding skills because descriptions load into every conversation.
-
MCP servers
- MCP connects agents to external services (build dashboards, logs, GitHub, Supabase) so the AI can read data and act outside the repo.
- Expo MCP exposes EAS builds, workflows, logs, and crash reports; install server tools for dashboard access and
npx expo install expo-mcpfor local capabilities (simulators, screenshots, logs). - Authenticate with
/mcp; update your start script to activate the local server. MCPs load all tool definitions up front and are token-heavy—install only when the project actually uses the service. - Use
expo doctorbefore debugging to catch version mismatches quickly.
-
AI workflow and context management
- Keep each conversation focused: one conversation → one outcome. Branch threads when moving from ideation to implementation to avoid context bloat.
- Monitor context usage (tools like
ccstatusline) and keep sessions below ~50% of the model window to reduce hallucination and cost. - When stuck on product direction, use ideation skills (e.g.,
/co-ceo) to settle core ideas before discussing tech choices.
Practical steps (concise)
- Scaffold:
bunx create-expo-app@latest habit-tracker && cd habit-tracker && bun install && bun start. - Add skills first (
npx skills add ...), then connect a single MCP if you need external logs or build control. - Validate fixes with the local MCP server (drives simulator, reads native/JS logs).
- Watch context: rename threads ("initial planning"), branch for stack decisions, and stop adding unrelated details to the same thread.
Outcome
Follow this pattern to let the AI act as a productive teammate: give it compact instruction files (skills), give it hands where needed (MCP), and manage conversation context so the model stays accurate and cost-effective.