AI & Prompt Engineering

An AI that can do more than answer one question at a time. You give it a goal, and it figures out the steps, takes them, checks the results, and keeps going until the job is done: without you holding its hand at each step.

A trick where you ask the AI to think out loud, walking through its reasoning step by step, before giving a final answer. It sounds simple, but it dramatically improves accuracy on math problems, puzzles, and anything that needs careful logic.

Markdown files kept at the root of a repo that describe conventions, commands, and non-negotiables for an AI coding assistant. The assistant loads them automatically, so they're where you put the rules you don't want to repeat in every prompt.

How much text an AI can hold in its head at once: both what you send it and what it writes back. A bigger context window means it can read longer documents, remember more of a conversation, or look at more files at the same time.

Showing the AI a few examples of exactly what you want before asking it to do the same for your real input. The examples teach it the pattern far more reliably than a long description ever could.

Taking an AI that already knows a lot and giving it extra training on a smaller, more focused set of examples. So it gets really good at one specific job or style. Unlike just writing a clever prompt, fine-tuning actually changes how the AI thinks.

When an AI confidently says something that sounds right but is actually made up. It happens because the AI is guessing what fits, not looking things up. The most dangerous hallucinations are the ones that are detailed and plausible.

A computer program trained on an enormous amount of text, books, websites, conversations, until it can read and write language the way a person does. Claude and ChatGPT are LLMs. They power chatbots, writing helpers, and tools that can work with code.

An open protocol that lets AI assistants connect to external tools and data sources — databases, issue trackers, filesystems, SaaS APIs — through a standard interface. An MCP server exposes a specific tool; the assistant can call it when you give permission.

Choosing which model in a family to run for a given task. Faster, cheaper models (like Haiku) handle narrow, well-defined edits; deeper models (like Sonnet or Opus) are worth their cost on ambiguous or multi-file work. Picking deliberately trades depth against speed and spend.

A mode in many AI coding assistants where the model can read your code and draft a step-by-step plan, but cannot edit files or run destructive commands until you approve. It's a cheap way to catch wrong assumptions before any code is written.

The message you send to an AI to get a response. A good prompt is specific, gives the AI enough background, and clearly says what you want back. Prompts are how you talk to an AI. And small changes to the wording can lead to very different results.

A way of giving an AI the right reference materials before asking it a question. The system searches a library of documents for the parts that look most relevant, hands them to the AI, and then asks the question. This way the answer is grounded in real, up-to-date information instead of the AI's general training.

Named, reusable packets of instructions the assistant can invoke on demand — often paired with a slash command. A skill bundles the steps, the context it needs, and any guardrails, so the same workflow runs the same way every time.

Shorthand commands you type with a leading slash to invoke a predefined skill or behavior in an AI assistant — `/commit`, `/review-pr`, `/plan`. They keep common workflows one keystroke away instead of one paragraph away.

Independent AI workers the main assistant can spawn to handle a scoped subtask — typically research, exploration, or parallel work. The subagent runs on its own context, does its job, and returns a short report, so your main thread stays focused on decisions.

A set of background instructions you give an AI before the actual conversation starts. It defines the AI's role, personality, and any rules it should always follow: like a job description it never forgets, no matter what the user asks next.

A dial you can set on most AI tools that controls how creative or how predictable the answers are. Low temperature means the AI plays it safe and gives the same answer every time. High temperature means it takes more chances and surprises you.

The little chunks of text an AI reads and writes in. A token is usually a short piece of a word: sometimes a whole common word, sometimes just a few letters. AI services bill you and limit how much you can send by counting tokens, not words.

When an AI is given access to outside tools, like searching the web, reading a file, or sending an email, and can decide on its own when to use them. Without tools, an AI can only talk. With tools, it can actually do things.

Asking the AI to do a task without giving it any examples first: just relying on what it already learned during training. It works well for common, everyday tasks the AI has seen many times before.