AI Features Need Product Edges, Not Just Better Prompts
Most AI features fail.
They do not fail because the model is bad. They fail because developers treat everything around the model like a demo.
I recently worked on an iOS workout app. It uses Claude for exercise plans. The model part is easy. You send context and get a plan back.
The hard work is what makes it a real product. You must handle:
• Monthly AI credit balances • Offline modes • Auth token storage • Loading states during generation • Different rules for adding vs swapping exercises • Tests for boring edge cases
This is where quality lives.
A button should not just call an endpoint. It must know if a suggestion is allowed. It needs to check if the user is online. It needs to check if they have credits left.
The model does not care about these things. The user cares.
When the AI responds, it must include the updated credit count. The UI must update immediately. This stops the user from seeing stale data. If the backend says they spent a credit, the UI must show it now.
I wrote 13 tests for one view model. I tested credit limits and offline errors. This is not for fun. This is because these things break in front of real users.
Connectivity failures must have clear messages. Do not say "unexpected server response." Say "You are offline. Reconnect to use AI features."
Shipping AI features is normal software engineering with a probabilistic dependency in the middle.
Focus on the contract around the model:
• Can the user use it right now? • What happens if the network dies? • Is usage tracked correctly? • Does the UI reflect the server state? • Can you test edge cases without the model?
Once you fix these, the feature stops feeling like a prompt on a button. It starts feeling like a part of your app.
Do not ask if the model answers. Ask if the feature survives real life.
Source: https://dev.to/toddssullivan/ai-features-need-product-edges-not-just-better-prompts-18k
Optional learning community: https://t.me/GyaanSetuAi