The words used across these notes in as plain English as possible.
Working with AI
Model
The part that actually produces the text. It has been trained on an enormous amount of writing and works by predicting what should come next. Claude and GPT are families of models.
Token
Roughly a word, or a piece of one. Models read and write in tokens, and most products charge by how many get used. Paying for every token means paying for the output whether it turned out to be useful or not.
Context
Everything the model can see at once: the instructions, the conversation so far, and any files that have been read in. It has a limit, which is why a long conversation eventually starts losing the thread.
Prompt
Whatever you send. A question, an instruction, or a whole document.
Session
One continuous conversation, sometimes called a chat. A new session begins with none of the previous one's context, which is the problem most of these notes circle around.
Agent
A model that has been given tools and permission to act rather than only reply. It can open files, run commands and keep going until a task is finished.
Fanning out
Splitting work across several agents running at the same time, in the hope of going faster.
MCP
Model Context Protocol. An agreed way of plugging an AI tool into another application so it can operate that application directly, rather than only describing what to do.
Usage allowance
The cap on how much of a product you can use before it stops, slows down or starts charging more.
The tools mentioned
Claude, ChatGPT
Chat applications. You type into a window and get an answer back.
Claude Code, Codex
Tools built on the same models, but able to open a project, change its files and run commands against it. The difference from a chat window is access, not intelligence.
Ways of working
Slice
One small part of a product, built all the way from the screen down to the stored data, so that it works rather than just looks finished.
Boundary
The written edges of a piece of work: what it covers, what it must leave alone, and what finished looks like.
Stop condition
The line in a task saying when to give up and come back with a question, rather than inventing an answer and carrying on.
Handoff
A short note describing where the work currently stands, written for whoever picks it up next.
Product notes
One document describing what the product is, what it has to remember, and which questions are still unanswered.
Dependency graph
A picture of what relies on what, used to work out the order things have to happen in and what breaks when one of them changes.
Programming words
Repository
The folder holding a project along with its full history of changes. Often shortened to repo.
Branch
A separate line of work, so changes can be made and tested without disturbing the version everyone else is using.
Revert
Undoing a change and returning the project to how it was before.
Data model
What the product stores and how those pieces relate. Getting it wrong early is expensive because everything built afterwards assumes it.
Schema
The agreed shape of some data. Which fields exist, what type each one holds, which are required.
Persist, storage
Saving something so it is still there after the program closes. A thing that does not persist disappears when the app does.
Stub
A placeholder with the right shape that does not do the job yet. Useful while building, a problem when it is mistaken for finished work.
Refactoring
Rearranging code without changing what it does, usually to make the next change easier.
Abstraction
A general-purpose layer added so several things can share the same code. Worth having once several things need it, and clutter before that.
Compiler
The step that turns written code into a program that runs. On a large project it can take minutes, which is why waiting on it shapes how people work.