I Built an AI Feature in a Week — Here's the Messy, Honest Truth

I used the Anthropic API with streaming responses. The API itself is straightforward — send a prompt, get tokens back. The 'AI work' was maybe two days: writing the prompt, handling the stream, parsing the output into the structure the UI needed. The other four days were: error handling, loading states, rate limiting, cost estimation, and making it feel fast.

The thing that surprised me most was how much the quality of the output depended on the structure of the prompt. A vague prompt produced vague summaries. A prompt with specific instructions about format, length, and what to prioritise produced output that felt like a professional had written it.

The hard parts aren't the AI parts. They're the same engineering challenges that make any feature hard: graceful degradation when the API is down, handling edge cases in the input, making the loading experience feel good when responses take 2-3 seconds, and keeping costs predictable at scale.

I added a simple character count check before every API call — if the document is under 300 characters, don't summarise, just show it as-is. That one guard eliminated about 15% of API calls and made the product feel smarter.