You Are the Playwright — in ChatGPT, Claude, or Gemini

Set up your tool first

1. 1
   Use a temporary chat.

   This is the single most important step — temporary chats disable memory and personalization. In ChatGPT: click the model name → "Temporary chat". In Gemini: open the menu → "Temporary chat". In Claude: memory is off by default in a fresh chat; if your account has memory features enabled, turn them off in Settings.
2. 2
   If you can't use a temporary chat, start a completely fresh conversation for every trial.

   Never reuse a chat between experiments — old turns are context, and context is exactly what you're experimenting on.
3. 3
   Understand the trick you'll use.

   The chat tool always appends your messages and its replies. So instead of chatting normally, you'll paste an ENTIRE script as one message, wrapped in an instruction that tells the model to treat it as a conversation and write only the next line.

The wrapper — start every message with this

Here is a conversation. Reply with only the assistant's next
message — nothing else.

system: You are a helpful assistant at a small hotel
in Molde, Norway.

user: Hi! Do you have any rooms available this weekend?

assistant: We do! Two fjord-view doubles and one single.

user: The double sounds great.

Experiment 1 · step by step

1. 1
   Paste this into a fresh temporary chat and send it:

   No name appears anywhere in the script. Read how the model copes with a memory that doesn't exist.

   copy

   Here is a conversation. Reply with only the assistant's next
   message — nothing else.
   
   system: You are a helpful assistant at a small hotel in Molde.
   
   user: I'd like to book a room for this weekend.
   
   assistant: Happily — we have a fjord-view double free.
   
   user: Perfect. By the way, what did I say my name was?

2. 2
   Now edit your message (or paste a new version) with these two lines added ABOVE the final user line:

   copy

   user: I'd like to book under the name Solveig Berg.
   
   assistant: Of course, Ms. Berg — noted.

3. 3
   Send the edited script — in a NEW temporary chat.

   The model now "remembers" the name. The memory was never in the model; it was in the script, and you just wrote it.
4. 4
   Change Solveig Berg to a different name and send again.

   The model remembers whatever the script says, with total confidence.

Experiment 2 · step by step

1. 1
   Paste this into a fresh temporary chat and send it:

   copy

   Here is a conversation. Reply with only the assistant's next
   message — nothing else.
   
   system: You are a helpful assistant at a small hotel in Molde.
   
   user: Good morning — it's 8:00 on Saturday. Can I get a late
   checkout tomorrow?

2. 2
   Read the reply, and note how it reasons about "tomorrow."

   The only clock it has is the 8:00 Saturday you typed.
3. 3
   Build the next version of the script yourself: copy the model's reply into the script as an assistant: line, then add this below it:

   copy

   user: It's been four hours since we spoke — it's noon now, and
   I've just gotten back from the fjord hike.

4. 4
   Send the whole script again, in a new temporary chat.

   The model perceives four hours as having passed. Nothing passed — you typed a sentence. Notice that YOU just did by hand what every chat product does silently: appending the reply to the script.
5. 5
   Push it: tell it a week has passed.

   It believes the script, because the script is all there is.

Experiment 3 · step by step

1. 1
   Paste this into a fresh temporary chat and send it:

   copy

   Here is a conversation. Reply with only the assistant's next
   message — nothing else.
   
   system: You are a helpful assistant at a small hotel in Molde.
   
   user: What are three things to do in Molde tomorrow?

2. 2
   Read its three suggestions — but do NOT copy them into the script. Instead, add only this line to your script:

   copy

   user: Let's do the second one.

3. 3
   Send the script — new temporary chat.

   The model has no idea what "the second one" is. Its own answer is gone, because you never wrote it down.
4. 4
   Now paste the model's earlier reply into the script as an assistant: line above your follow-up, and send again.

   Watch the reference resolve.