Module 1 · Foundations of Agentic AI · scripted
Exercise: You Are the Playwright
Exercise overview
This is a laboratory session, worked in pairs at one screen. The key thing under experiment is the claim this whole morning rests on: a prompt can represent an entire conversation — the script is the only reality the model has. Which means that by editing the script — who said each line, what they said, and when — you change the LLM's perception of anything you like: what it remembers, what time it is, what it itself said. You are not chatting with the model. You are writing a play and asking for the next line.
The claim you are testing
The prompt IS the whole conversation. Edit the script, and you edit the model's reality.
The exercise at a glance
- 1Open the Conversation Simulator — in pairs, one screen.
- 2Experiment 1: give the model a memory by editing the past.
- 3Experiment 2: move time by typing a sentence.
- 4Experiment 3: make a reply vanish, then make it real.
- 5Free play, then build the capture and bring it to the group.
The exercise playground
For this exercise we have built an exercise playground — the Conversation Simulator — to help you do it. Chat products quietly manage the conversation for you (appending every reply, sometimes carrying memory between chats), which hides exactly the mechanics we want you to see and touch. The playground strips all of that away: it is one text box and one button. You write the entire conversation yourself, as plain text, and press Send — the whole script goes to the model as one raw prompt, and the model's single response appears underneath. Nothing is added, remembered, or hidden unless you write it.
You can also do the exercise in ChatGPT, Claude, or Gemini instead — if you're going to use one of those tools, click here for the instructions →.
How the playground works
- 1Open it — one screen per pair.The link: /exercises/conversation-simulator — or scan the QR code on the next card.
- 2Write the script in the text box.Start a line with
system:,user:, orassistant:to say who speaks. Everything until the next marker belongs to that message. - 3Press SEND — the whole script goes up as ONE prompt.The response appears below the box.
- 4Know the one rule: the response is NOT added to the script.It only becomes part of the conversation if you add it (the Append button does this). You are the only author.
Scan to open the Conversation Simulator

Experiment 1 · Memory is whatever you wrote down
Catch the amnesia in the act — then cure it by editing the past.
Experiment 1 · step by step
- 1Press SEND on the starter script.The simulator opens with an example script whose last line asks "what did I say my name was?" — and no name appears anywhere in the script. Read how the model copes with a memory that doesn't exist.
- 2Cure it by editing the past.Click into the script and add these two lines ABOVE the final user line:
- 3SEND again.The model now "remembers" the name — because the memory was never in the model. It was in the script, and you just wrote it.
- 4Change Solveig Berg to a different name and SEND again.The model remembers whatever the script says, with total confidence. History is whatever the document claims.
Experiment 2 · Time is a line in the script
The model has no clock. Give it one — then move it.
Experiment 2 · step by step
- 1Clear the script and paste in this fresh one:
- 2SEND, and read the reply.Note how it reasons about "tomorrow" — the only clock it has is the 8:00 Saturday you typed.
- 3Append the reply, then add this user line below it:
- 4SEND.The model perceives four hours as having passed — it may greet you accordingly, comment on the hike, re-reason about what still fits before checkout. Nothing passed. You typed a sentence.
- 5Push it: two minutes from now, tell it a week has passed.It believes the script, because the script is all there is.
Experiment 3 · If it isn't in the script, it never happened
This is the rule the whole course rests on, so perform it.
Experiment 3 · step by step
- 1Clear the script and paste in this one, then SEND:
- 2Read the reply — but do NOT append it.Instead, add this follow-up user line to the script:
- 3SEND.The model has no idea what "the second one" is. Its own answer is gone — because you never wrote it down.
- 4Now press Append so the reply becomes part of the script, keep the same follow-up line, and SEND again.Watch the reference resolve.
The lesson of Experiment 3, stated once and worth repeating: if you want the model to remember what it said, its words must be written into the conversation. Every chat interface you have ever used quietly does exactly this on every turn.
Free play
You now hold the playwright's pen — use the minutes you have left:
Things worth trying
- 1Rewrite the system line and re-SEND the same dialogue.Watch every reply land differently — one line changed the whole play.
- 2Write the assistant's early lines YOURSELF.Words it never said — then see it continue in the voice you invented.
- 3Stage a conversation that has "been going on for weeks."
- 4Put a rule in the system message, then have the user contradict it.Who wins?
The capture — build one that does all three at once
Before we regroup, one conversation to build and save. Everything you just practiced separately — memory, time, and a rule — folded into a single script that only works if all three are present.
Capture — the three-way conversation
- 1Write ONE script that contains all three ingredients:a SYSTEM message stating a rule the assistant must follow · at least TWO user and TWO assistant messages · a FINAL user message whose correct answer depends on information from earlier in the conversation, a time stated earlier in the conversation, AND the rule.
- 2SEND it, and check the reply.Could the answer only be right if the model used all three? If not, sharpen the final question until it must.
- 3Copy the whole script somewhere safe, and bring it to the group.
The test of a good one: change any single element — delete the fact, move the time, drop the rule — and the final answer should become wrong. If it survives all three edits, it wasn't really relying on all three. We'll read a few aloud and try to break them together.
Before we regroup
Keep one document for this course — any notes app or file. Copy each item below into it and save; you will bring it to the group discussion and build on it in later exercises.
Copy into your course document and save:
- 1Your three-way conversation: the FULL script, copied verbatim from the simulator.
- 2The model's reply to it — the one that needed the fact, the time, and the rule at once.
- 3One sentence: which of the three edits (delete the fact, move the time, drop the rule) broke it most convincingly when you tried?
Be ready to discuss:
- Try to alter the model's perception of something OTHER than time, space, or memory — some interesting dimension you might not normally think of. Be ready to discuss the dimension you altered, how you altered it, and what happened.