how we do it
Notes
Short how-tos for the decisions that come up when you are making something with an AI agent. Read them in order the first time, or jump to the one you need.
01
start tiny
Finish something small
Make something small enough to finish today. Test it yourself. Put it on a link a friend can open.
02
name it
What kind of project
“App” means several different things. Say website, phone app, computer app, or a program that just runs.
03
pick one
Pick an agent
Cursor, Claude Code, or Codex. Chat apps are for talking. The terminal is for building. Auto is fine until it isn’t.
04
starting point
Leave a note for the agent
Drop an AGENTS.md in your project. Start from ours. Rewrite it until it sounds like you.
05
type here
The terminal
A text window for talking to the computer. On a Mac, use iTerm2. Here are the commands you will actually need.
06
two tools
Git and GitHub
Git saves snapshots on your computer. GitHub holds a copy on the internet. You need both if you want a backup and a live site that updates.
07
one set
Which tools to use
Do not shop for tools. Pick one set, finish the project, then get bored on purpose.
08
optional
Your own web address
A vercel.app link is already live. If you want a custom name, buy it in Vercel so DNS is not a second project.
09
ask first
Ask for options
Do not lock your first idea in. Ask what the agent thinks, get options, pick, then steer.
10
disposable
Start many chats
A new chat per idea. Do the thing, commit, push, check, iterate, close. Then start fresh.
11
you think
Do the thinking yourself
If you are just clicking next, yep, go, you are not learning anything or doing anything valuable.
12
not broken
Not feeling broken
Not feeling broken matters more than heaps of features. Build on something that works.
13
handy tip
A thousand tabs
Close the tabs you're finished with. A thousand open tabs is usually just poor organisation.
14
handy tip
Good emails
Put the point in the subject. A first email has to stand alone. A reply can be one sentence.
15
use sparingly
Follow-up messages
A message sent while the agent is still working waits in a queue. Send one when it matters. Otherwise wait.
16
you decide
Don't keep unshipped backwards compatibility
Keep the old way only if someone is already using it. If the feature never shipped, tell the agent to delete it.
17
before work
Say it back
Ask it to say the idea back. Ask what it thinks. If it calls the idea too big, tell it to explore.
18
the path
User flows
A flow is a path with steps. Write the real ones down. Vague vibes are not flows.
19
1 to 5
Scoring user flows
Easy = can they get through. Good = does the path do the job. Two scores, one sentence why each.