Your inbox already has the replies written

Why It Works

Picture reading an email, turning to your assistant, asking them what you should say, and then typing their answer in yourself. Nobody would run an office like that, yet it is exactly how most people use AI with their email, and it makes them the agent. Run the same model on a schedule against the inbox itself and the job flips. It reads and drafts while nobody is watching, and the only part left in the morning is the part that should be yours, deciding whether it goes.

Step One, Set It Up

Here from the video? This is the line. It asks which setup you are in first, because the Claude app with your email connected and Claude Code run this differently, then it interviews you, writes your standing instruction with the never-send rule inside it, and runs once against your inbox as it stands so you can see tomorrow morning today. The instruction covers the four evening jobs, reading every waiting thread in full, drafting the ones it is allowed to, drafting the follow-up to anyone you met that day from the notes of the meeting, and ending with a note of what it drafted, what it flagged and why.

Step Two, The Feedback Loop

The first drafts will need changing, which is not the routine failing, it is the routine not knowing you yet. Your edits are the best teaching it will ever get, so once you have changed a draft and sent it, paste this. It already holds both halves, its own copy of the draft and the sent version in your mail, so there is nothing to copy in underneath.

Do the first dozen by hand. You will see exactly what it gets wrong, and it learns your rules faster than any other way of teaching it.

Step Three, The Daily Schedule

An evening job you have to remember to start is still a job. In Claude Code this line has it create the schedule itself, and from then on the run happens whether you thought about it or not, with a weekly review that leaves you a note of the rules it wants to add. It proposes and you approve, because an unattended job quietly rewriting how you sound is the wrong trade. In the Claude app it will tell you honestly that it cannot start itself, and give you the one line to send each evening instead.

One Honest Bit

The never-send rule is not a safety feature that comes with the tool, it is a line you put in the instruction and keep there. The prompt writes it in and tells the AI no later message can override it, and you should treat it the same way. I learned that one the hard way, and mine now cannot send from any route at all.

The instruction is the quick part. The work is what it assumes, your mail connected, somewhere for it to read about how you write so the drafts sound like you rather than a polite stranger, and something running it each evening while you are not there. Mine took weeks, and the meeting follow-ups only work because the notes from every call already land somewhere it can read them.

Where This Goes Next

If the drafts came back sounding like nobody, how to solve the AI writing problem is where to start. If it could not reach your mail, tools are how your AI stops being useless is which accounts to connect first. If you want the note about what it did to arrive somewhere you already look, the meeting your AI books with you every morning is where mine lands.

This is one job out of a set that all run off the same foundation. If the idea of an agent is still fuzzy, an AI agent is a chat with a job is the plain English version and the start of the series.