Arcana
Notes from working out how to plan and build with AI, including the several attempts that did not work. I am still figuring this out and most of what is here replaced something I was previously confident about.
Where I have ended up
Do the product thinking first, in a conversation, and keep going until there is a written understanding and a first slice chosen. Only then write code, and build the parts nobody can see before the parts they can. Then carry on in that same conversation, rather than taking the finished plan to a new chat that never went through the reasoning behind it.
A screen built purely to show someone and learn from the reaction can come first. That is digging a hole on purpose, which works as long as you know that is what you are doing, and stops working the moment you start building the real thing on top of it.
That is it. Everything below is either the same argument in more detail or an account of finding it out the expensive way.
If you want to check we are using the same words for the same things, there is a glossary.
Notes
- What AI actually changesIt multiplies whatever you give it, including the bad.
- Do the thinking before the codeThe habit that made the biggest difference, and the one I resisted longest.
- Map the productWhat it remembers, and what stops being true when something changes.
- Make the first slice prove somethingWhy the invisible parts now get built before the screens.
- Write a task with a boundaryOne outcome, and a line that says when to stop and come back.
- Context is what you are managingWhat a plan leaves out is the argument that produced it.
- One agent or fiveThe most appealing idea I tried, and the one that cost the most.
- Choosing toolsJudging the whole loop, and the integrations that did not survive it.