Notes on building the AI-native web.
Page 16 of 17
Write the evaluation before you build the AI feature
If you can’t describe what a good answer looks like, you can’t ship the feature — you can only demo it.
How many dependencies is too many?
Every package saves a day now and costs a little forever. A test for whether the trade is worth making.
The knowledge base you need before hiring an AI employee
Most companies discover their documentation is a wiki nobody trusts. That discovery is the project, and it pays for itself either way.
Do you actually need to fine-tune a model?
Almost certainly not yet. Three cheaper things solve the problem you think fine-tuning solves, and they’re reversible.
Do author pages help you get cited?
Attribution is a credibility signal, and most company blogs publish everything under the company name. What that costs.
Do you need a separate page for every question?
No — and the sites that tried it built a thin-content problem instead of an authority one. How to decide what deserves its own URL.
Why software estimates are wrong, and what to do instead
Not because engineers are optimistic. Because an estimate is a forecast of work nobody has fully specified yet.
Getting the tone right
Most AI teammates sound like a press release. The fix is examples, deletions, and resisting the urge to make it charming.
Images are the slowest thing on your site
And the usual fix — an on-the-fly transformation service — trades a size problem for a latency and billing problem.
Notifications: the feature that annoys everyone
Every stakeholder wants their event to notify. The result is a system users mute, which then fails to reach them when it matters.
Build vs buy in the AI era: the questions that decide it
Building got cheaper, which changed the maths but not the principle. Five questions that settle the argument faster than a spreadsheet.
Meeting transcription: what to do with all that text
Recording every meeting is easy now. Most companies accumulate thousands of transcripts nobody reads and call it knowledge management.