Print this page or copy the boxes into a note. Checking a box doesn’t save anything anywhere. Use the boxes to spot what you can already explain and what you want to review.
Quick terms
- Token: A small chunk of text that the model reads and counts. Sometimes it’s a whole word, and often it’s a piece of one. AI providers list their prices per token.
- Context: Everything the model sees when you send a message. That includes your text, any earlier messages the app sends along, and anything else the app adds.
- Retrieval: Having the app fetch the right document and add it to what the model sees.
- Agent: An AI set up to use tools that take actions in other systems, like issuing a refund.
- Fine-tuning: Training a model further on your own examples to change how it behaves.
- Use / Watch / Ignore: Our three tool verdicts. Use means we recommend it for a named job. Watch means there isn’t enough public info yet. Ignore means it doesn’t fit the job.
Before you explain how a model works
- In one sentence, I can say what the model actually sees when I send a message: my text, earlier messages, and anything the app adds.
- I can explain what a token is without just calling it a word or confusing it with a password.
- When an app seems to remember, I know which thing is happening. Either it re-sends our earlier chat with each new message, or it saves notes somewhere.
- I can find the current token rates on the provider’s pricing page, and I write down the date I checked.
Before you pick an approach
- If the job needs a specific document, I’ll have the app fetch it, or I’ll paste the relevant section in, before I consider training a model.
- If the job means changing something in another system, like issuing a refund, I’ve named the tool that does it and who’s allowed to use it.
- If the job is about tone or format, I’ve already tried putting a few examples in the prompt, or I’ve written down why I can’t.
- I’ll test any improvement on new examples that weren’t already in the prompt or the training data.
Before you repeat one of our tool verdicts
- I’m looking at the tool’s current profile page, not going from memory.
- I opened the live pricing page and wrote down the date. Any price in my notes came from that page.
- ChatGPT: https://openai.com/chatgpt/pricing/
- Claude: https://claude.com/pricing
- Copilot: https://github.com/features/copilot/plans
- Cursor: https://cursor.com/docs/models-and-pricing
- The Ollama app uses the MIT license, which is open source. The model I’d download has its own license, and I’ve opened its model card, the info page that comes with the model.
- I can explain why none of the five tools got an Ignore.
After the lab
- For each lab job (A, B, and C), I wrote down my chosen approach, what result would prove that choice wrong, and what I’d avoid doing first.
- I can explain why I picked a prompt, document retrieval, an agent, or fine-tuning for each job.
The lab jobs are the handbook that changes, the refund that has to hit billing, and the short replies that drift off the approved templates. The three approaches you choose among are a plain prompt, fetching a document, an agent with a named tool, or more training. A price you didn’t just read on the vendor’s page doesn’t belong in the note.
Sources & limits: The tool verdicts come from our profiles as checked on 2026-11-07. Prices change, so always use the live pricing pages.