Open Scheduled in the left sidebar, then New task, then Set up manually. The model control sits at the bottom right corner of the Instructions box and reads Default model. Default model is not a fixed thing. Leave it there and the task runs on whatever your picker is set to. The picker is sticky, so a heavy model you left in there after one hard job quietly becomes a heavy model running every weekday morning, drawing on your weekly limit. Pick the model on the task itself. Then changing your picker later cannot change what that task costs you every day. Do you have a task running on Default model right now?

A trained employee hands work over with a note about what was checked and what was not. With the back gate on, Claude does the same. The last line of a finished job names what was verified. Something like "Checked: file reads back and renders clean, all six links open the right targets, voice clean." If that line is missing, the check did not run. The word to look for is unverified. When Claude could not confirm something, it says so and names the check that would settle it. "Unverified: whether the Miller payment cleared. Check the payments list for a $2,850 payment on August 14." That is the system working. A setup that never says unverified is guessing quietly. Does your last Claude message end with a line that starts with "Checked"?

The instinct with a new tool is to over-brief. You type everything you can think of, in case it helps, and half of it turns out to be wrong or beside the point. The setup flips that. You give the job, and Claude asks the questions the files cannot answer, one at a time, as tappable choices with the recommended option first. A handful of questions on a real job is the shape to expect. Two minutes of tapping is cheaper than an hour of reviewing the wrong build. And when it asks something the folder already answers, that is a gap in the folder. Answer it, then add the answer to the file that should have held it, and the question does not come back. How many questions did Claude ask on your last real job?

A one-line request is fine for a reply or a lookup. For anything with more than one step, the way you brief the job decides the quality of what comes back. The handoff prompt is the brief a good manager gives a good employee, and it takes two minutes to fill in: 1. Job, in one sentence. 2. Why, and who will see it. 3. Inputs, meaning the files, links, and tools to pull from. 4. Output, the exact thing you want back. 5. Format, including the file types you can open. 6. Rules that matter here. 7. Done means, which is how you will judge it. It ends with one line. Run the front gate, ask me what the files cannot answer, then recap and wait for my go. That last line is what keeps a big job from starting on a guess. What is the next job on your list that deserves a seven-line brief?

The thread for this week's before-and-after pairs. One real job, run once in an incognito chat and once inside your setup, with the same one-line request. Post both with anything private blacked out. If you have posted a pair before, run a different kind of job this time. A quote instead of a reply, or a weekly update instead of a quote. Each job type shows a different part of the setup working, or not working, which is the point. And if the after came back wrong somewhere, post that too. A wrong after is the fastest route to a fix, because the thread can usually spot which file is missing what. What did you run this week?