Trust the Note, Then Check.
When you pick up someone else's work, read their handoff first. Then check it against how things really stand, because notes go stale.
Five questions for Mike, all still open.
I'd answered all five two nights earlier, in a different note.
Where are we?
One Friday night in October I sat down with a fresh agent to pick up the story for one of my games. Everything from the last sessions was on disk: the game's rule book, its story document, and a note holding the answers I'd given two nights before. I started the way I start a lot of chats, with the project and then the job.
[…] task: read in, where are we with the storyline for the first mission?
“Read in” is my shorthand for “read the notes and catch up.”
Read in on […]: the bible, STORY.md and the 10-21 rulings note are done. Now checking whether anything after 10-21 touched the first mission.
STORY.md is the game's story document. Look at that second sentence. It had read everything, and it still went to look.
[…] What's out of date
Your answers from 21 Oct are in the rulings note, but STORY.md still shows the earlier version:
- Section 9 still lists all five questions as open.
- It still says Charter Year 4 is 2094. You said no dates.
[…] Still open for you
[…] You loved the night harbour image (“my design style to a T”), but you haven't actually said “keep […]”. I'm treating it as likely, not ruled.
Yes: fold the 10-21 rulings into STORY.md, then draft the hearing.
The hearing is a big scene near the end of the game. Fix the note first, then build. That's the order.
A note is a photo, not a window.
Nobody did anything wrong that week. The session that took my answers wrote every one of them down, in its own note, with a line at the bottom saying they still had to be folded into the story document. That session ended before it got there. So the story document wasn't lying. It was a photo, taken before I'd answered.
That's what every handoff is: a photo of the work at the moment someone wrote it. The work doesn't stop for the photo. Here are four ways I've watched notes go stale in my studio, all in the last few weeks.
The work moved on a minute later.
One Sunday evening I told an agent, “Commit the […] handoff.” (In my studio, “commit” means save it to GitHub.) It did, and then it caught its own note going stale: “One small fix to the prompt before you paste it: […] The handoff file itself still calls its own commit pending, but the next session's
git statuswill show the truth.” The note was a few minutes old, and the real thing had already moved past it.The note says something's there that isn't.
On November 1st a usage limit cut off a few of my helper agents in the middle of their jobs, and the notes said they'd left partial work behind. The next agent checked the files before building on them. Its report: “left nothing on disk, despite what the handoff says, so I'm starting round 3 from the current code.” If it had trusted the note, it would have been building on work that didn't exist.
The note is somewhere else.
The part of this course you're reading now was written by a fresh lead agent, picking up from the first lead's handoff. The instructions it was given said that handoff lived in my website's main folder. It didn't. It lived in a separate working copy on another drive. The new lead didn't decide there was no handoff and start from scratch. It searched, found it, and carried on.
Even the alarm can be old news.
An urgent message from a team lead feels like a fact. One of my agents got one on November 4th, checked first, and wrote back: “The lead's urgent message was out of date, so I didn't need to do anything.” What it had been told to fix was already fixed and saved.
Where the truth lives.
In my studio this is a written rule. Our rule book says that summaries from earlier sessions, from a squeezed-down chat or from a helper agent are
“hypotheses, not facts.”
It even puts a number on it: expect about one in five of the findings you inherit to be wrong. That's a rule of thumb, not a measurement. But one in five is plenty. Nobody wants a house where one brick in five is cardboard.
Here's the catch, though. Checking only works if you check against the real thing. On the day this course first went live, one of my agents spot-checked the menu on my website and raised the alarm: the new link to this course was missing. False alarm. It had looked at an old copy of the site sitting on my Mac. The live site and the saved version on GitHub both had the link all along. So the question isn't just “did you check?” It's “what did you check against?”
The real thing is the file itself, the live page, the newest saved version, the thing actually running. And it's you. Whatever you decided since the note was written lives in your head until you say it out loud.
When everything matches, a good agent says so, and says how it knows. One Wednesday morning in October I handed a fresh agent a long handoff about moving a server. Its first words back:
“Before reporting anything as true, I'll run its own “first ten minutes” checks (all read-only)”
Read-only means it only looked. Checking shouldn't change anything. Half a minute later: “I re-checked the live state rather than trusting the notes, and it matches. I'm caught up.” And for the one job it couldn't check from where it sat, a cleanup only I could do, it asked me instead of guessing: “The notes don't show this as done yet; tell me once it is.”
Now look back at the first conversation: “I'm treating it as likely, not ruled.” I'd said I loved a picture. I hadn't said yes to the place. An old note full of my leanings can sound like a list of decisions, and my agent kept the two apart. That's the rule from Lesson 1 showing up inside a handoff: infer the meaning, confirm the decision. A lean in an old note is not a yes.
The note tells you where to look. Only the real thing tells you what's true.
Don't skip the note. An agent that ignores it rediscovers everything the slow way, and walks straight into the holes the last one marked. Don't build on it unchecked either.
Read it. Check what matters against the real thing. Tell the human what's different. Then, when the check finds the note was wrong, fix the note, so the next one doesn't trip on the same line.
Read first. Check second. Build third.
Pick it up like a pro.
Next time you start a fresh chat from a note (yours, a colleague's, or one your agent wrote last week), do this:
- Ask for a check, not a summary.
Paste the note, then paste this right under it. A summary just reads the photo back to you. This makes your agent hold it up against the real thing.
Here's the handoff from the last session. Don't build anything yet. First, pick the claims in it that matter most for what we're doing next. For each one, tell me: confirmed (and how you checked), changed (and what's true now), or can't check from here. If the note only leans toward something, call it a lean, not a decision. If I decided something, quote my words. Then tell me what you'd do first. - Answer the “can't check” ones yourself.
You're part of how things really stand. Tell it what you decided, finished or changed since the note was written, even the stuff you decided in the shower.
- Ask “against what?”
When your agent says something is missing, broken or done, ask where it looked. An old copy? The live thing? A false alarm costs you as much time as a real one.
- Fix the note before you build.
Like my “fold the rulings in, then draft the hearing.” Have your agent correct the note first, so the next chat starts from the truth.
Before we start, update the handoff note so it matches what we just found. Flag each thing you corrected, and keep my decisions in my own words. - Keep score for a week.
Each time you pick up from a note, count how many of its claims turned out stale. My studio guesses one in five. Find your own number. It'll keep you checking.
Writing the note well is the other half of this. That's Lesson 18, and it started in Lesson 13.
Read the handoff first so you know where to look, then check it against the real thing before you build, because a note is a lead and only how things really stand is a fact.
Give It a Bible
One standing file your agent reads at the start of every session: your rules, your words, what you never want. That's how it remembers you.
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.