Building Products with AI in 2026

After building several AI products — including 拾光 Glimmer and various internal tools — here’s what I’ve learned about shipping AI features.

Start with the User Problem

It’s easy to get excited about AI capabilities and build technology looking for a problem. Don’t.

Start with:

1. What’s the user’s pain point?
2. How are they solving it today?
3. How can AI make this 10x better?

Manage Expectations

AI isn’t magic. Set clear expectations:

• What the AI can and can’t do
• When it might fail
• How users can provide feedback

Build for Failure

AI outputs are probabilistic. Your product needs to handle:

• Unexpected outputs
• Slow responses
• Service outages

Always have a graceful degradation path.

The Build Loop

1. Ship something simple
2. Watch how users interact
3. Improve based on real usage
4. Repeat

Don’t try to build the perfect AI feature. Ship, learn, iterate.

Cost Awareness

AI API calls add up fast. Track costs from day one:

• Cost per user
• Cost per feature
• Cost per successful outcome

Build in caching and optimization early.

What’s Next

The tools are getting better every month. The constraint isn’t technology — it’s imagination and execution.