Code5302

How ChatGPT and Claude work

Links

Guide

How ChatGPT and Claude work: one-page cheat sheet + 5 prompts

A short glossary of the 7 concepts from the carousel, 5 ready prompts that turn each one into a habit, and the official sources. Numbers come from OpenAI's and Anthropic's own pages (checked October 22, 2026).

Cheat sheet

ConceptWhat it isWhat it means for you
LLM (large language model)A model trained on a huge amount of text that generates text. ChatGPT and Claude are built on one.It's not a search engine. It's a very good prediction machine.
TokenThe unit the model reads: a word, part of a word, or a single character. In English, 1 token ≈ 4 characters and 100 tokens ≈ 75 words (OpenAI). Other languages differ.Limits, prices and context are all counted in tokens. Repeating yourself burns them.
Next-word predictionIn pretraining, the model learns to predict what comes next (Anthropic). It still plans ahead: in a poem it picks the rhyme before writing the line.It continues whatever you start. Give the role, the format and one example up front.
Context windowThe model's working memory: all the text it can see right now. Every question and answer piles up. Up to 1M tokens on Claude Opus 5.5.Claude lists "current conversation length" as a factor in your usage limit. New topic, new chat.
TemperatureThe randomness dial. High = creative and varied, low = the most likely answer. Anthropic: even at 0, outputs aren't fully identical.Ask important questions 2-3 times; disagreement is a red flag.
Knowledge cutoffThe date the model's knowledge stops. Claude Opus 5.5: reliable knowledge through June 2026.For prices, versions, laws or news, turn on web search or paste the page.
Web search / RAGThe model pulling in documents or search results while answering. That's how it reaches anything outside its training data.Want current, sourced answers? Give it the document or turn on search.
HallucinationA confident answer that's false. OpenAI: tests reward guessing and "I don't know" scores zero. Anthropic: Claude's default is to not answer, and a familiar-sounding name can switch that off.Permission to say "I'm not sure" plus a source request cuts errors. It doesn't end them.
System promptThe hidden instruction the app maker puts at the start of every chat. ChatGPT's Custom Instructions do the same job for you.Tell it who you are and how you want answers, once.

5 ready prompts

1. Fewer made-up answers: cite sources, admit uncertainty

Put this at the top of your message.

When you answer, say "I'm not sure" anywhere you're not confident.
Never invent a number, name, date or link.
Next to every claim, give the source and the exact sentence from it.
If you can't find a source for a claim, delete the claim and write [no source].

My question: [YOUR QUESTION]

2. Save your limit: move a long chat to a fresh one

Run this at the end of a long chat, then paste the result into a new chat.

I'm moving this conversation to a new chat. So I don't have to explain everything again:
1. State my goal in one sentence.
2. List the decisions we've made.
3. List the questions that are still open.
4. List every number, name and rule you'll need in the new chat.
Keep it under 15 bullets. Don't repeat the conversation, just the outcome.

3. Catch outdated info: check if it's current

For anything that changes: prices, versions, laws, news.

My question: [YOUR QUESTION]

Before you answer:
1. Tell me up to what date your knowledge on this is reliable.
2. Could the answer have changed since then? If yes, why?
3. If it could have changed, search the web and answer with a dated, linked source.
If you can't search, say so and tell me which official page to check.

4. Catch inconsistencies: compare answers

Ask the same question in 2-3 separate chats, then paste the answers into a new one.

Below are 3 answers I got to the same question.
1. List what all three agree on.
2. Point out every place they contradict each other.
3. For each contradiction, which one is more likely right, and why? Say so if you're not sure.
4. List the claims I should verify and where to verify them.

Question: [QUESTION]
Answer 1: [...]
Answer 2: [...]
Answer 3: [...]

5. Your own system prompt: Custom Instructions

ChatGPT: Settings → Personalization → Custom Instructions. Claude: Settings → profile instructions. Write it once; it applies to every chat. Fill in the brackets.

I'm a [your job / what you do]. I mostly ask about [what you use it for].
When you answer:
- Give the short answer first, then explain if needed.
- Write plainly, no filler intros.
- Say "I'm not sure" when you're not; never invent numbers or dates.
- For anything that changes (prices, versions, laws), say how current your info is.
- If you need more info from me, ask instead of guessing.

Sources

Note: these techniques reduce made-up answers; they don't eliminate them. For anything that matters, check the original source.

Looking for a different code?

YoodaAI logo

YoodaAI

Career AI

Free to start

Everything you need to land the job: practice the interview, apply at scale, and get live help while you talk.

  • Interview practice

    Mock interviews with instant feedback

  • Bulk job apply

    Send hundreds of applications in one go

  • Real-time interview copilot

    Live answer hints while you are in the call

  • Free CV builder

    ATS-ready templates at no cost

yoodaai.proTry it free
AIAutoFlow logo

AIAutoFlow

AI automation

For businesses

Custom AI automation for companies. Scale your operations with intelligent workflows built specifically for your business.

aiautoflow.proExplore solutions