Most bad AI output starts with a one-line brief. Someone types "write me a client update," gets back something an intern could have written, and decides the tool is overrated. Nobody would hand a new hire a sticky note and walk away. People do it to AI all day.
I run most of my business with Claude now. Last month it rebuilt my website and set up the email sequence behind it. When it hands me something useless, I've stopped blaming the model and started reading my own brief, because that's where it went wrong nearly every time. If you're still comparing ChatGPT against Claude to find the smarter one, you're testing the wrong thing.
The file it reads first
Every project I run with Claude has one document at the top. Its first line reads "Read this first in any new chat." Under that: what's live, what's half-built, what I've parked, what I still owe, how I like to work. When I open a new chat I don't re-explain my business. I point at the file.
It's the binder you'd hand a good new hire on day one, and it isn't free. I update it at the end of every working session, because a stale binder is worse than none. It still drifts. Last week it said my blog had three placeholder posts. There were four. It said scorecard signups were tagged by stage. They weren't; the stage lives in a field. Claude caught both by checking the real system instead of trusting my notes.
A new hire who checks the binder against reality is one you keep.
Last week's brief
I wanted the blog and newsletter live. The brief was one goal line, five steps in order, and one sentence at the end:
Finish the blog page.
Stop newsletter signups from landing in my scorecard follow-up sequence.
Add a signup box to the blog.
Build the newsletter template.
Draft the first post.
Ask me before publishing anything or changing the live journey.
The order was the part that mattered. My email automation fires on anyone who joins the list. If the signup box had gone up before step 2, someone who asked for a weekly post would have landed in a sales sequence instead. Nothing in "set up a newsletter" tells you that. You know it because you've seen the plumbing, so you say it.
The last sentence did the rest. I didn't hover. Claude worked through the setup and stopped exactly twice, once before it touched the live automation and once before it deleted anything. Those were the only two places I wanted a say.
Rules I only want to say once
I don't use em dashes. I don't want motivational sign-offs, and I keep a list of words that make writing sound generated. For a while I fixed those by hand in every draft. Now they're written down where every chat can see them.
Facts about me go in the same place. An early draft said I "directed engineers" at Microsoft and Meta. I didn't. I ran programs that had engineers on them, and I led teams of program managers. Anyone who has done either job knows those aren't the same thing. I corrected it once and it hasn't come back.
Two businesses, two binders
I also run an AI music education business. It has its own project, its own history and its own status file, and nothing crosses over. That's what you'd want from a person splitting time between two teams.
Even with all of that, Claude opened the wrong website on that same job. I have three sites on one Squarespace account, and I'd asked it to "finish the blog page." It noticed, backed out without changing anything, and I sent one line: make sure you're on joshuagilliland.com.
That one was on me. Three keys on the ring, and I didn't say which door.
Talk first
A lot of my briefs start as voice memos. I ramble, half the thoughts don't finish, and I end with "pick that apart." That only works because the file already carries the background. The memo just has to carry what's new.
Next time an answer disappoints you, read your prompt back as the new hire who received it. If you wouldn't know what done looks like, it didn't either.
