How to choose the right home for your AI workflow

Executive overview

AI tools default to building dashboards and apps because they're trained on code — but that's rarely what you need. Most workflows belong inside the AI workspace itself, not in a deployed application.

There are three places any AI workflow can live: the workspace (projects/skills), a viewable artifact or shared file, and a full application. Defaulting to the simplest option saves time, cost, and maintenance.

The right question is not "can AI build this?" but "where should this live?"

The three locations

• Workspace (projects or skills): handles 85–90% of use cases; simplest, lowest friction
• Viewable output (live artifact or shared HTML file): right for recurring data views shared with others
• Full application: valid only when external users need to log in and can't share a folder or AI instance

Projects vs. skills

• A project is a focused workspace primed for one topic + one task (e.g., ACME client → weekly follow-up emails)
• A skill is portable across conversations; useful for things like brand guidelines applied to any presentation
• Projects scale freely; skills should stay under ~15 and must be distinct — the AI picks skills by title/description and will choose the wrong one if they overlap
• Prime a project with a system prompt (instructions) and optional knowledge files (context, past examples)

Viewable outputs: artifacts and shared files

• Live artifacts (Claude desktop): support live data connectors (CRM, calendar, email); update in real time; not shareable — one user only
• Shared HTML file: ask AI to produce a standalone HTML file, drop it in a synced folder (OneDrive, Google Drive, Dropbox), share with team; update by feeding new data to Claude and regenerating
• The shared HTML approach has no hosting, no login, no deployment, no security overhead

When to build a full application

• Only if external users need to log in and don't share a folder or AI subscription with you
• Hidden costs: hosting, database, security, user auth
• API cost trap: AI baked into a project or skill runs on your subscription; AI in a deployed app bills per token via API key — usage from multiple users can get expensive fast

The decision test

Answer in order; stop at the first yes:

1. Is this a recurring, focused activity? → Build a project
2. Is this task applicable across multiple topics/conversations? → Build a skill
3. Does this output need to be viewed — by you only? → Use a live artifact
4. Does it need to be viewed by others? → Shared HTML file in a synced folder
5. Do external users need to log in, outside your shared folders/AI instances? → Build an application