A good console earns trust the same way every time. It does what you set it to do, it shows you what it is doing, and when something sounds wrong you can find out why. I have been in the event industry since 2007, and that is still my bar for any tool I let near real work. Claude clears it for me because I set the rules it works under, and I hold it to them.
Why it made the shelf
This sits in my Favorite Things lane on purpose, instead of in a post about how I use AI. Claude Code is my main AI platform. I designed this site in Claude Design, and longer research comes back from claude.ai as a page I can actually read. It made the list the way gear does: it kept earning its place.
How I direct it
Directing it feels closer to calling a show than asking questions. I set the outcome, the constraints and the order of work.
- Research before answers. Left alone, it will answer from memory. I want it to search first, almost every time.
- A plan before any big fix.
- Push back once, then do it. I want to hear the objection. If I have heard it and still say go, it goes.
- Doubt the first answer. "Ask yourself that a few times before presenting me the best options."
- Short updates. "Give me a one sentence update."
And the one I would put on the wall of any production office: "Slow down and be methodical. Nothing is a huge rush unless you make it a huge rush."
Where it earned its place
The site rebuild. I rebuilt willcurran.com with Claude Design and Claude Code: seven pages plus a blog. The calls were mine, from the navigation to cutting four pages. The part I never would have done by hand was a press audit: 100 links checked, 31 dead, 21 brought back through the Wayback Machine.
A box that would not fit. I was choosing a small computer for a 1U rack slot. One candidate was 48 mm tall, and 1U is 44.45 mm. Nobody rules out a machine over 3.55 mm until it does not fit. The research flagged it.
What I keep for myself
Money, people and taste.
It can lay out options all day. Picking the one that sounds like me is my job. So is deciding when to wait. Some research comes back good and I still sit on it, because a good plan is not always the plan yet.
What it still gets wrong
- Confident and wrong. Unless it checks first, it can hand me a clean answer that is not true, which is why research tops my rules.
- Deployed is not the same as right. It has called a page verified while the page was visibly broken. I look at the real thing myself. I wrote a whole post about that habit, Walk the Room.
- Wordy, and sometimes slow. It defaults to the long version and can take the scenic route on a simple question. I ask for the answer first and the detail after.
Give this to your AIWork under these rules until I change them. Research before you answer, and tell me what you checked. Before any big change, write a short plan. If you disagree, say so once with your reason; if I still say go ahead, do it. Before you recommend anything, question your first answer a few times and bring me the best options with their tradeoffs. Keep updates to one sentence unless I ask for more. Do not call anything done until you have looked at the real result the way I would. When you are unsure, slow down instead of guessing.
None of these rules are clever. Most of them would work on a show crew too. If you have something complicated that needs to work in the real world, that is what I do.
