Your AI doesn't remember yesterday, at least not the part that matters. A new chat can see how things look now. It can't see why you made them that way. So it helpfully puts the old way back.
The fix is a short note. After every real fix, write four lines: what changed, why, how we checked it, and what's still open. Keep the note where the AI reads it first. This guide has the template, a filled-in example, where to put it in Claude and ChatGPT, what never goes in it, and prompts that write the note for you.
Why the AI keeps undoing your fixes
Every new chat is the AI's first day on the job. Because it is. Claude Code's docs put it plainly: "Each Claude Code session begins with a fresh context window." (A session is one stretch of work. The context window is everything the AI can see while it works.) And Claude's help page on projects says context isn't shared across chats in a project unless you add it to the project's knowledge.
So say you changed how bookings get confirmed. Tomorrow's chat sees the new setup, not the reason. It looks a little odd, so the AI "fixes" it and puts the old way back. You explained the reason once, in some chat, somewhere. Good luck finding it.
Honestly, this one's mostly for me. I bounce between projects all day. The note is how I stop explaining the whole business again.
The note doesn't retrain the AI or change it for good. It's information the AI reads at the start, like a sticky note on the monitor for whoever sits down next.
The four lines
After every real fix, write these four lines. A real fix is one you'd hate to see undone. One or two sentences each, plain words, date on top.
| Line | What to write | Why it's there |
|---|---|---|
| What changed | The change, in plain words, and where it lives. | So the AI knows the new way is on purpose. |
| Why | The problem it fixed, or the decision behind it. | The one line the AI can't figure out on its own. |
| How we checked it | The test you ran and what happened. Or "Not checked yet." | So nobody trusts a fix that was never tested. |
| What's still open | What's left to do, or a question nobody has answered. | So the next chat starts where you stopped. |
HANDOFF NOTE · [date]
What changed:
Why:
How we checked it:
What's still open:Where to keep it so the AI reads it first
Put the note where the AI reads it at the start of every chat, not buried in an old chat. Newest note on top.
| If you use | Put the note in | How |
|---|---|---|
| Claude (the app) | A project's instructions | Click Projects (or go to claude.ai/projects), then + New Project. In the project, click Set project instructions, paste the note, and click Save instructions. Claude uses them in every chat in that project. Free accounts get projects too, up to five. |
| ChatGPT | A project's instructions | Select New project in the sidebar. In the project, open the more options menu (•••), select Project settings, and add the note to the instructions. They apply only inside that project. |
| Claude Code | The CLAUDE.md file in your project folder | A plain text file. Claude Code reads it at the start of every session. |
| OpenAI Codex | The AGENTS.md file in your project folder | Codex reads AGENTS.md files before doing any work. |
| Anything else | The first message of a new chat | Keep the note in a doc. Paste it in first, before you ask for anything. |
What never goes in the note
- Passwords, API keys and login codes. (An API key is a password one app uses to talk to another.) Write where it's kept, like "the key is in the password manager," never the key itself.
- Customer details. No names, phone numbers, addresses, or anything about one customer's job. Write "a customer," or use a test name.
- Payment details. No card or bank numbers, ever.
- Whole chat transcripts. Paste the decision, not the conversation. The four lines are the point.
If you share a project, the people you share it with can see its instructions. Write the note like a coworker will read it. They might.
When to update it
- After every real fix. Not every tiny edit. Just the ones you'd hate to see undone.
- When the plan changes. Fix or delete the old note that same day. Otherwise tomorrow's AI gets very confident instructions from a version of the project you already gave up on.
- When something "still open" gets done. Write a new note. Its "What changed" line closes the old question.
- When the notes get long. Keep the newest up top and move the rest to a notes file. Short wins: Claude Code's docs say the more specific and concise your instructions, the more consistently Claude follows them.
Prompts: let the AI write the note
You don't have to write the note from memory. At the end of a session, ask the AI to draft it, then check it yourself. At the start of the next one, tell the AI to read it first. These work in Claude, ChatGPT, or any AI assistant.
Before we stop, write a handoff note for this session, so the next chat can pick up without me explaining everything again.
Use exactly these four lines, one or two plain sentences each, with today's date on top:
What changed:
Why:
How we checked it:
What's still open:
Rules:
- Only include what we actually did and decided in this chat. If we didn't test something, write "Not checked yet." Don't guess.
- "Why" matters most. Give the real reason for the change, in my words if I said it.
- No passwords, API keys, login codes, or customer names, numbers or addresses. If one matters, say where it's kept instead.
- If we made more than one real fix, write one note per fix.Before you change anything, read the handoff notes in this project's instructions (or in CLAUDE.md or AGENTS.md if we're working in a code folder). Then tell me in three short bullets: what we changed most recently, why, and what's still open. If anything I ask for today would undo one of those changes, stop and ask me first.Here are my handoff notes. Find any that conflict with each other or look out of date, list them, and tell me why. Don't delete or rewrite anything until I say so. Then suggest a shorter version that keeps every decision that's still true, newest first.
[paste your notes]Read the draft before you save it. If the "Why" line is vague, fix it. That's the line you'll need most.
Copy this: the note checklist
BEFORE YOU SAVE THE NOTE
[ ] Date on top
[ ] What changed, in plain words
[ ] Why: the real reason
[ ] How we checked it
(or "Not checked yet")
[ ] What's still open
[ ] No passwords, keys or codes
[ ] No customer details
[ ] Saved where the AI reads
it first
[ ] Old notes it replaces are
deleted or updatedExample (made up)
Everything here is made up: a demo project and a test customer.
Northside Heating & Air uses an AI assistant to help run its booking messages. Customers used to get a booking confirmation by text and by email. Some thought they'd booked twice and called to cancel. So the team switched to text only, and wrote this note:
HANDOFF NOTE · 2026-10-02 (DEMO)
What changed: Booking
confirmations go by text only.
The email copy is turned off.
Why: Customers got a text AND an
email and thought they'd booked
twice. Some called to cancel
the "extra" visit.
How we checked it: Booked a test
visit for "TEST Jordan Sample"
on the office test phone. One
text came. No email.
What's still open: Reschedule
messages still send an email.
Switch those too? Decide Friday.The next day, someone asks the AI to "make sure every customer gets a confirmation email." With the note in the project's instructions, the AI can see the email was turned off on purpose, and why. With prompt 2 above, it's told to stop and ask before undoing it. Without the note, all it sees is a missing email.
Mistakes to avoid
- Skipping the "why." The change is easy to see. The reason isn't. That's the line that stops the undo.
- Pasting the whole chat. The AI has to dig for the decision. Four lines beat a wall of chat.
- Leaving old notes up. A note for a plan you dropped is worse than no note. Delete it or mark it replaced.
- Saying it's tested when it isn't. "Not checked yet" is a fine answer. A guess isn't.
- Putting a password in "just for now." Write where it's kept instead.
- Treating the note like a lock. It's information, not a rule the AI can't break. Claude Code's docs say it treats these files "as context, not enforced configuration." Still check what the AI changes.
Stop explaining it twice
Four lines after every real fix: what changed, why, how we checked it, what's still open. Keep them where the AI reads first, keep them current, and keep secrets out. I already explain enough things twice. I'm not paying a computer to forget the same thing I forgot.
Checked on Oct 4, 2026 against
Every claim about a third-party tool in this guide (plans, prices, menu paths, commands, limits) was checked against these official pages on Oct 4, 2026. These screens change often: if something looks different, trust the page over this guide.
- Claude Help Center: How can I create and manage projects?Projects for all users, free accounts up to five; Projects or claude.ai/projects → + New Project; Set project instructions → Save instructions, used "for all the chats within the project"; the + button for project knowledge; context isn't shared across chats in a project unless it's added to project knowledge; viewers of a shared project can see its instructions.
- Claude Help Center: Use Claude's chat search and memoryMemory is on by default for Free, Pro and Max; Claude saves memory on its own as you chat; managed in Settings → Memory.
- OpenAI Help Center: Projects in ChatGPTNew project in the sidebar; more options (•••) → Project settings for instructions, which apply only within that project; add files or paste text; save a reply with Save to project (label may vary).
- OpenAI Help Center: Memory in ChatGPT"Memory does not retain every detail from every conversation."
- Claude Code docs: How Claude remembers your project"Each Claude Code session begins with a fresh context window"; CLAUDE.md is read at the start of every session; treated "as context, not enforced configuration"; specific, concise instructions are followed more consistently.
- OpenAI Codex docs: Custom instructions with AGENTS.md"Codex reads AGENTS.md files before doing any work."
