A Letter to a Stranger.
The next agent knows nothing. Write the note it needs to pick up your work cold.
“Fixed the thing. Finish the rest tomorrow.”
What's done and how we know. What's running. What's next. Where it all lives. What you decided, in your words.
I don't want to be surprised again.
One Sunday afternoon my studio hit its usage limit with no warning. My lead agent, the one that keeps all my other agents on track, stopped answering me in the middle of a conversation with a one-line notice: come back this evening. Underneath it, at least one job had stopped halfway and nobody noticed for a while.
Once I got things going again, I told it what I wanted from then on.
Sounds good! Update working notes from each subsession and update your master handoff for the next Studio Maestro now and update with each new milestone as they occur. We still have some time, but I dont want to be surprised again.
“Studio Maestro” is what I call that lead agent. “Subsessions” are the agents working under it.
Done. The notes are current, and the routine is set so you won't be surprised again.
Every session got your order word for word. Each updates its own handoff now, with today's answers folded in, and again at every milestone: each job finished, each launch, each gate. […]
My master handoff is rewritten for the next Studio Maestro, then saved to GitHub with the shared pointer updated. It covers:
- what's paused and running;
- every answer you gave today, word for word;
- the usage rules you set;
- the questions held until Tuesday;
- a short, ordered next-steps list.
I'll rewrite it at every milestone from here.
↑ Read that list again. In five lines, it's most of this lesson.
The next agent is a stranger. A smart one.
In Lesson 13 I said to write your handoff like a letter to a friend. Keep the warmth. But the friend you're writing to has never met you. It doesn't know what “the thing” is, which file you mean, or that you already tried the obvious fix.
It is smart, though. You don't explain how to do the work. You explain where the work is. My studio's rule book says every note is written for a cold reader. Here's what that reader needs, in order.
What's done, and how you know.
“Done” on its own is a claim. “Done, and I opened the live page on my phone and it worked” is a fact. Write down how each thing was checked, and say plainly which things weren't. (Lesson 4 is all about why.)
What's still running, or waiting.
A job that's halfway, a reply you're waiting on, something paused on purpose. A stranger who doesn't know it's running might start it again, or kill it. Say what, where, when it should finish, and what to do if it doesn't.
What's next, in order.
Not a wish list. Short and numbered, most important first. Both of my agent's adjectives, short and ordered, are doing work.
Exactly where everything lives.
The real file name, folder and link. The account's name, never its password. “The draft” means something tonight and nothing tomorrow, even to you.
Every decision, in the human's own words.
So important it gets its own section, below.
Two more earn a spot when the work is big. How to do it again: the agent that built the first lessons of this course left a whole section on how they were made, so its successor could make more. You're reading one. And what went wrong, which Lesson 13 already covered.
Last thing: newest update on top, dated, saying whether it replaces anything older.
Write down what I said, not what you think I meant.
Look back at that list: every answer, word for word. Not a summary of my answers. My answers.
A summary is one agent's reading of what I meant. If the reading was a little off, the next agent inherits the mistake and can't see it. My studio keeps one file of every call I make, and each entry has the same shape: the time, my exact words, then separately how the agent read them and what it did. Here's a real one, about handoffs:
“[…] I will need you to comprehensively update your handoff and provide for me, please, a prompt here that I can copy and paste into the next session, as I have multiple cloud accounts […]”
“Cloud = Claude”
“Cloud” was supposed to be Claude, the AI. I fixed it in three words, and the record kept both, the slip and the fix, instead of tidying them into its own sentence.
That's the course's spine again. I talk loose and the agent works out what I meant. But what I decided stays in my words, so any agent, any day, can check its reading against the source.
Some things aren't tasks. They're promises.
A to-do list ends. Some things are meant to go on: a daily reminder, a permission you gave, a house rule you made today. They live in the agent's head, and that's the part that doesn't come along. One Thursday morning I'd asked my lead agent to remind me every day to send a report by email.
[…] One honest limit: these reminders live in this session and expire after seven days. So I've written it into my handoff and memory as a standing item, and every maestro who takes over renews it. You won't lose it.
Yeah, as long as you write it in the magic handoff, we remain what is it? We have contingency of government.
Ha! Continuity of government, exactly. 🏛️ It's written into the magic handoff now, under a heading every successor must carry: […]
Even if I compacted mid-sentence, the next me would pick up every one of them.
“Contingency of government.” I meant continuity. It knew.
Under that heading it listed seven things: the reminder, a permission I'd given it that morning, a few house rules, names we'd settled on that day. None were jobs to finish. All were things a stranger would quietly drop.
So give promises their own heading, near the top: carry these forward, every time. Each new agent renews them and passes them on. That's continuity of government. Or contingency, if you dictate like me.
Write it as you go, not at the end.
The end doesn't always announce itself. A limit hits, a laptop dies, a chat fills up faster than you thought (Lesson 17). The handoff you planned to write at the end never gets written.
So update it at every milestone. Whenever the end comes, the letter's already in the envelope.
Write a letter to a stranger.
Pick something you're working on with your AI that will take more than one sitting. Then:
- Have your agent fill in this template.
Paste it near the end of a session. Save the result where you'll find it, under the same name every time.
Write a handoff note using exactly the outline below. Write it for a stranger: someone smart who has never seen this chat and knows nothing about me or this work. Use exact names, file names, folders and links, never "the file" or "the thing". Quote my decisions word for word, then add your reading of them separately. If you don't know something, say so instead of guessing. Never include passwords or keys. HANDOFF: [what this is], written [date and time] Newest update goes on top. Say whether it replaces anything below. 1. WHAT THIS IS The goal in one or two sentences, and who it's for. 2. WHERE IT STANDS - Done: [item]. Checked by: [how we know it works] - Done but NOT checked: [item] - Still running or waiting: [what, where, when it should finish, what to do if it doesn't] 3. NEXT, IN ORDER 1. [most important first] 2. 3. 4. DECISIONS, IN MY OWN WORDS - [date and time]: "[my exact words]". Read as: [your reading]. Done: [what we did] 5. STANDING PROMISES (carry these forward every time) - [reminders, permissions, house rules, anything that must outlive this chat] 6. WHERE EVERYTHING LIVES - [exact file names, folders, links, account names] 7. HOW TO DO IT AGAIN - [the steps or tools that made today's work, so it can be repeated] 8. WHAT WENT WRONG - [mistakes, dead ends and false alarms, so nobody repeats them] 9. QUESTIONS WAITING ON ME - [most important first, one per line] - Give it to a stranger before you need to.
Open a brand-new chat and paste in only the note, nothing else. Then paste this:
You're picking up this work cold, and this note is all you have. Don't start any work. List every question you'd need answered before you could carry on, and point out anything vague, like "the file" or "fixed it", where you can't tell exactly what's meant.Every question it asks is a hole in your note. Take them back to your first chat and have it fill them in.
- Check the decisions section yourself.
Is it really what you said? If the agent tidied your words, ask for them back.
- Update it at every milestone.
Each time something finishes or you make a call, say “update the handoff.” Then a surprise stop costs minutes, not a day.
Write every handoff for a smart stranger, keep your decisions in your own words and your promises under their own heading, and keep it current as you go, because the end won't wait for you.
Trust the Note, Then Check
When you pick up someone else's work, read the handoff first, then check it against how things really stand. Notes go stale.
Every exchange on this page really happened and is quoted exactly, slips included; […] marks where I trimmed. Times are Eastern. My agent here is Claude, made by Anthropic.