Glossary 2 — Talking to Models: Prompts & Context

Glossary 2 — Talking to Models: Prompts & Context

AI Universe City · The AI Agency Learning Community · Glossary entry 2 of 8 · Human-supervised · 2026-09-16

By the end, you can: write a prompt like a briefing instead of a wish, and explain why the model "forgot" you.


1. Prompt

Plain: Everything you send the model at once — your question, your rules, and any text you paste.

Analogy: The work order you hand a contractor. A vague order gets a vague house.

Where again: Prompting 101 and every hands-on module.

2. Prompt engineering

Plain: The skill of writing prompts that work well each time. A good prompt sets a role, gives examples, and fixes the format and rules.

Analogy: A clear email to customer service. Same problem, far better answer.

Where again: The Build With AI track. A cheap skill with a big payoff.

3. Context window

Plain: How much text the model can use at once — your rules, the chat so far, and your files. Past the limit, the oldest text is gone.

Analogy: A whiteboard with a fixed size. To add new notes, you erase the old ones.

Where again: Lesson 3, where we count it in tokens. Also every "why did it forget?" moment.

4. System prompt

Plain: A hidden note that sets the model's role and rules before the chat starts. Most users never see it.

Analogy: The new-hire handbook. Staff hand it over before the first client walks in.

Where again: Building assistants — every agent lesson.

5. Conversation memory (chat history)

Plain: Each turn, the app sends the old messages back with your new one. That is continuity, not the model "remembering you."

Analogy: A barista who re-reads the full chat tape before each reply.

Where again: Privacy lessons; debugging new chats that start blank.

6. Temperature

Plain: A dial for how bold the model gets. Low keeps its picks close to the top choice. High reaches for rare words, so it can surprise you.

Analogy: Order "the usual," or say "chef, surprise me."

Where again: API settings. Voice agents run low. Brainstorming runs high.

7. Few-shot prompting

Plain: You put two or three worked examples in the prompt. Then the model copies your plan.

Analogy: "Write my toast — here are two I liked." The examples speak for themselves.

Where again: The fastest fix when the output "isn't the format."

8. Grounding / RAG (retrieval-augmented generation)

Plain: First, look up real docs. Then hand the model the right parts. Now it answers from your facts, not its own memory.

Analogy: An open-book exam. Same student, far fewer made-up answers.

Where again: The Professional track. Our main guard against hallucination (L1).


Check yourself: You paste a 90-page manual. Three follow-ups later, it answers from what it already knows. What fell off the whiteboard? (The context window — oldest content drops out.)

Artifact — post below: One before/after prompt pair for a real task (before = wish, after = briefing). Name which term fixed it.

Next: Lesson 3 — Running AI: tokens, TPS, VRAM & RAM.