Painful Lessons from Early AI Projects
If you're just jumping in, you may find these notes helpful.
I've been building a website and a finance and valuation module with AI. These are a few things I've learned along the way.
Set the expectations you would at work.
By assistant, I mean the AI I'm speaking with in the chat. I started treating it like a member of the team. At work, I would look for an answer before asking someone else to find it. I expect the same here: check the project documents you have permission to access, try to resolve the question, and tell me what you checked before asking for help.
Turn corrections into rules.
If I have to explain the same thing twice, I ask the assistant to record a rule for the project. Where to look for documents. Which methodology governs the work. What needs my approval before publication. Then I check whether the next response follows it. Writing down a rule is not the same as following it.
Describe the screen in terms a machine can check.
I kept saying there was too much space on our mobile pages. The heading took up most of the first screen; the content came later. A screenshot helped, but so did being literal: how much of the screen is the banner using? Where does the first useful content begin? How large is the empty area? Ask for a phone-sized screenshot of the result. Smaller text is not a fix for oversized margins.
Explain what cannot get lost.
For the valuation module, I wanted simpler screens without losing financial detail. Show the chart or table through a button, with the period, calculation and source available when needed. I was asking for less clutter, not less information. That distinction needs to be in the instructions.
Give the review a reference.
Find a way to measure the quality of the responses within the conversation. Provide an example and explain what matters about it: the writing, the layout, the calculations or the sourcing. Ask where the draft falls short against those criteria. Save that response and use it to assess the revision. A score out of ten is not useful if the standard changes between drafts.
Ask what another agent will contribute.
An agent can be another AI instance assigned part of the work, such as checking calculations or testing a page. It may use the same underlying model as the assistant. Calling it a specialist does not establish expertise. I want to know what it checked, what evidence it used and where its findings differ from another review. Agreement alone does not settle the question.
Keep the requests and the results together.
I ask for dated folders containing the available transcript, source documents, deliverables and test results. A summary is not a raw transcript, and an incomplete record should be labeled. Keeping the material together lets me return to the original request, compare revisions and take the same evidence to another AI without starting over.
By Aref M. Bajwa | September 16, 2026