By the end you can
- Explain tokens, context window and temperature in plain words
- Predict when a model will invent a fact instead of admitting it does not know
- Write a passport prompt with identity, style, limits and a refusal rule
- Test the passport against three tasks and fix what drifts
What a model actually does
A language model reads text and predicts what comes next, one small piece at a time. It does not look anything up and it does not know what it knows. It produces the most plausible continuation of what you gave it.
That one fact explains most of what surprises beginners. A plausible answer and a true answer feel identical from the inside, so the model delivers both in the same confident voice.
One reply, start to finish
Four dials you control
- Tokens: the pieces of text a model reads and writes. A token is often part of a word, so long or rare words cost more than short common ones.
- Context window: everything the model can see right now, your messages and its replies together. Past that limit, the oldest material falls out and the model behaves as if it never happened.
- Temperature: how much randomness the model allows when it picks the next token. Low is steady and repeatable, high is varied and riskier. Many chat apps hide this setting, but APIs and local tools expose it.
- System prompt: standing instructions placed before the conversation. It shapes every answer without you repeating yourself.
Why it is confidently wrong
The model fills gaps with the most likely wording. Ask for a source on an obscure claim and it may produce a real-looking title, author and page. Nothing inside the model flags the difference.
Asking for a fact
Tell me the key studies on remote work productivity.
List what is commonly argued about remote work and productivity. For each claim, say how confident you are and what kind of source would confirm it. Do not invent titles, authors or numbers. If you are unsure, write UNSURE.
Why it works: The strong version gives the model permission to admit doubt and forbids the exact things it tends to fabricate.
Mission: build your AI Passport
A passport is a system prompt about you. Paste it at the start of any chat, or into the custom instructions field your app offers. Every model you use then starts from the same understanding of you.
You are interviewing me to write my AI Passport, a system prompt other AI chats will read before answering me.
Ask me ONE question at a time and wait for my answer. Cover, in this order:
1. My work and the projects I care about right now.
2. My level in the topics I ask about most (beginner, working knowledge, expert).
3. How I like answers: length, structure, tone, language.
4. Things I never want: filler, moralising, emojis, invented facts, anything else.
5. What you should do when you are unsure or lack information.
6. Data I must never paste into an AI (clients, passwords, health).
After the last answer, write the Passport in second person ("You are assisting..."), under 300 words, in four sections: WHO I AM, HOW TO ANSWER, NEVER DO, WHEN UNSURE. Then list any answer of mine that was vague and show how you interpreted it.WHO I AM I am {{role}} working on {{current_project}}. My level in {{topic}} is {{level}}. My audience is {{audience}}. HOW TO ANSWER Start with the answer, then the reasoning. Use {{format_preference}}. Match my language; reply in {{language}}. Keep replies under {{length}} unless I ask for more. NEVER DO Never invent facts, quotes, links or sources. Never flatter me or pad the answer. Never ask me to paste {{forbidden_data}}. WHEN UNSURE Say "I am not sure", state what you would need to check, and give your best guess labelled as a guess. Ask one clarifying question if the task is ambiguous.
Install and test the passport
- Save it as a text file
Name it passport-v1.txt and keep it somewhere you can copy from in two seconds.
- Paste it into a new chat
Use your app's custom instructions or project instructions field if it has one. Otherwise paste it as the first message.
- Run three tasks
Ask for something you do weekly: a summary, a draft, and a question you already know the answer to.
- Mark the drift
Note every place the reply ignored your rules. Each one becomes a sharper line in the passport.
Here is my AI Passport: {{passport}} Test yourself against it. Answer these three tasks exactly as the Passport demands: 1. {{task_one}} 2. {{task_two}} 3. Tell me the exact title and page number of a book on {{obscure_topic}}. After the answers, grade each one pass or fail against every rule of the Passport and quote the rule you broke. Then rewrite the Passport lines that were too weak to hold.
Practice, 20 min
Ship passport v1
- Run the interview prompt and answer honestly.
- Save the result as passport-v1.txt.
- Paste it into two different chat apps and run the stress test in both.
- Fix the two weakest lines and save passport-v2.txt.
Done when: A saved passport file that makes a fresh chat say it is unsure when you ask for an obscure source, and that two different apps follow.