The AI terms that actually matter for getting work done

Executive overview

Most AI jargon is noise. A small set of terms directly affects how productively you use these tools. This video cuts through seven categories — tools, interaction methods, outputs, customisation, data retrieval, techniques, and technical foundations — keeping only what changes how you work.

The terms worth knowing are the ones that change what you can do, not what you can say.

AI tools and providers

• AI providers give access to chat interfaces: ChatGPT, Claude, Grok, Gemini.
• AI browsers embed AI throughout the browsing experience, not just in a chat box (e.g. Comet from Perplexity, Atlas from OpenAI).
• Coding agents (Claude Code, Cursor, Codex) wrap a model with scaffolding specifically for writing code — the agent is the tooling, not just the model.

How to interact with AI

• Dictation converts speech to text — faster for getting complex thoughts into the AI quickly.
• Voice-to-voice / advanced voice mode lets the AI speak back; useful on commutes or walks.
• Canvas (OpenAI/Gemini/Grok) and artifacts (Claude) are collaborative editing surfaces — highlight text, request targeted edits, or generate interactive code visualisations.
• Agent mode lets the AI take autonomous actions in a browser, including clicking, writing, and sending on your behalf.

Outputs

• Reasoning models think before responding; you can see the reasoning trace and catch wrong directions early.
• Deep research sends the AI off for 10–20 minutes to produce a 20–40 page report on a specific question.
• Image generation tools (Midjourney, Flux AI, Imagen, Imagen 3/Nano Banana) now support editing yourself or objects into existing images — useful for ads and content.
• Video generation (Sora 2, Veo 3.1) supports adding a personal cameo into generated video.
• Structured outputs / JSON are machine-readable formats the AI produces when it expects its output to feed into another app — ignore if you're the end reader.

Tailoring AI to your workflow

• System prompts / instructions are what the AI reads before any user message — they lock it to a specific role or task.
• Projects are folders with a system prompt and optional files, enabling repeatable tasks (e.g. auto-drafting a recurring email from a transcript).
• Custom GPTs (OpenAI-specific) are shareable projects distributed via the GPT app store.
• Claude skills bring a project's capability into an existing conversation without switching context, then step away when done.
• MCP (model context protocol) is the backbone that lets connectors link AI to external apps like Gmail or Slack, enabling AI to read, draft, and send without manual copy-pasting.

How AI retrieves data

• RAG (retrieval-augmented generation) pulls a specific chunk from a large external document into the model's memory to answer a question — a one-shot lookup.
• Semantic search finds information by meaning, not keywords — matches intent rather than exact terms.
• Agentic search runs a continuous loop using multiple tools to explore files and folders until it finds what's needed — higher accuracy, takes longer, used heavily in coding agents.
• Hallucination is the AI producing false or fabricated information; grounding is the fix — either citing sources in the prompt or enabling web search with citations.

Prompting and coding techniques

• Prompt engineering is writing clear, structured, specific requests — strong communicators outperform technical users here.
• Context engineering goes further: shaping all the data the AI pulls in (files, web results, code) not just the text you type — the quality of the surrounding context determines output quality.
• Vibe coding spans a spectrum from full trust (accept everything, don't read the code) to high scrutiny (review every line). As models improve, the balance shifts toward more trust.

Technical foundations

• Context window is the model's working memory, measured in tokens (sub-word units). ChatGPT supports ~400k tokens; Gemini 2.5 Pro ~1 million. More context degrades intelligence — keep it lean.
• Training is the process of teaching a model on large datasets; practically relevant because you may want to opt out of your data being used.
• Agents run a model in a loop with tools, making autonomous decisions until a task is complete.
• Sub-agents (popularised by Claude Code) get their own isolated context window for a subtask, keeping the main thread lean and the overall output quality high.