Lexicon

What an AI keeps hold of about you and your project after the current conversation ends.

Not the same as the context window, which is only what fits in front of it right now. Memory is the part somebody decided was worth bringing back.

Working with an AI that goes off and does a whole job in a loop — looks things up, uses tools, checks its own work — instead of answering one question at a time.

The grown-up version of vibe coding. The difference is the checking. An AI that cannot check its own work is just a very fast typist.

Machine-made output that is long, generic, and worth very little.

Merriam-Webster made slop its 2025 Word of the Year, the same year Collins picked vibe coding, which tells you most of what you need to know about that year. It covers code, documents and bug reports now, not just writing.

How far a change can reach: every file and every behaviour it could possibly touch.

Say it out loud in your prompt and the AI stays in its lane. Leave it unsaid and a one-line fix comes back wrapped in four hundred lines of tidying you never asked for.

Code that already exists and already works, which you now have to change without breaking it.

This is most real work, and it is the hard kind. Most vibe-coding advice quietly pretends otherwise.

Copying what something looks like without understanding why it works.

From a talk Richard Feynman gave in 1974. The modern version is a project full of settings files copied off a blog post, none of which anyone can explain.

An AI working through the steps out loud before it gives you an answer.

It gets better answers on anything with several steps in it. It is not a recording of what the machine actually did inside, and reading it that way will fool you.

A test that writes down what code does today, right or wrong, so you can tell later whether you broke it.

The one thing to do before letting an AI near old code. Pin down what it does now, or you will not be able to tell a fix from a fresh break.

An answer that reads well, looks tidy, sounds certain, and is wrong.

The signature failure of this whole field. People ramble and hedge when they are unsure; machines do not. So sounding sure stopped being proof of anything.

The part of your prompt that says what the AI is not allowed to do: which tools, which files, which lines it must not cross.

The most useful part of any prompt, and the part people skip. An AI will not guess your privacy rules. If it is not in the constraints, it is not in the result.

Deciding on purpose what goes in front of the AI: the instructions, the files, the notes, the tools, the earlier conversation.

It replaced prompt engineering as the phrase people use, and the shift matters. The leverage moved from how you word the question to what you put in front of it — and just as much, what you keep out.

What happens when you hand an AI so much material at once that the answers get worse.

More stuff is not more understanding. Past a point, the thing it needed is buried in everything else you added. The usual fix is a second, smaller AI with one narrow job.

When something you said early in a long session quietly stops counting.

Also called context decay. You spot it when the AI breaks a rule you set forty messages ago, with no sign it ever heard it.

How much the AI can hold in front of it at one time.

Bigger is not free. It costs more, runs slower, and makes the part you actually needed harder to find. The skill is picking what goes in, not piling more in.

The moment a prototype that demoed beautifully turns out to be impossible to build on.

Everything before the cliff was generated. Nothing before it was designed. You usually find out when the second feature has to share something with the first.

Five fill-in-the-blank sentence shapes that turn a vague request into something you can actually test.

Written by Alistair Mavin and colleagues at Rolls-Royce and published in 2009, long before anyone was prompting anything. The five shapes cover: always, when this happens, while in this state, if this option is switched on, and if something goes wrong. It is the cheapest upgrade most vibe coaders can make.

A way of turning text into numbers, so a computer can tell which pieces mean roughly the same thing.

The machinery under search that understands what you meant instead of matching the words you typed. Similar is not the same as useful, which is why a second sorting pass usually pays for itself.

Writing the check before the feature, so you can tell whether a change actually made anything better.

The same habit as writing the test first, adapted for a machine that never gives the same answer twice. Without it, “the new prompt seems better” is all you have, and that is not evidence.

Repeatable checks that score what an AI gave you against what it was supposed to give you.

Your safety net for something that will not answer the same way twice. Growing up means trading “that looked right” for a number you can watch over time.

Training an existing AI a bit further on your own examples, so it behaves the way you want.

Usually the wrong first move. Most problems people bring to fine-tuning turn out to be about what you showed it or how you asked, and both are far cheaper to fix.

A brand new project with nothing in it yet, where the AI can write whatever it likes.

This is where the demos come from. It is also the easy half, which is why a greenfield demo impresses nobody who has actually shipped anything.

When an AI confidently makes something up: a command, a quote, a setting, a fact that does not exist.

The word flatters the machine. It suggests it saw something wrong, when really it just wrote something that sounded likely. Sometimes likely is false.

Everything built around an AI to turn it into something that can do jobs: standing instructions, tools, retries, memory, and the wiring between them.

Also called scaffolding. These days most of the improvement on long jobs comes from a better harness, not a smarter model, which is why builders talk about it more than they talk about the model itself.

The one point where a named person looks at the work and takes responsibility for it going out.

If you cannot point at yours, you do not have one. It is a person and a moment, not a document.

The gap between what you asked for and what the AI reasonably assumed you meant.

“Add login” leaves out almost everything, and the AI fills the gap with its own defaults, which are rarely yours. Drift is not the machine being wrong. It is the machine being sensible about something you never said.

The heart of vibe coding: ask, look at what came back, correct it, ask again, until it matches what you meant.

The whole craft lives here. Everybody shows you the first go. Almost nobody publishes the fourth, which is where you find out whether you understood the problem.

Model Context Protocol: an agreed way to plug an AI into outside tools and data.

Anthropic put it out in late 2024 and it became the standard plug. A tool that does not speak it is effectively off the grid.

Getting all your notes, rules and reference material together before you type the first prompt.

Borrowed from restaurant kitchens, where it means everything is on the board before the burner goes on. The disaster it prevents is the same in both jobs: making it up halfway through because you did not prep.

Sending different jobs to different AIs: the cheap fast one for the boring parts, the expensive one for the hard thinking.

What a good team lead does with people, applied to machines. Once you route, you stop having a favourite model and start having a way of working.

One person doing what a small team used to do: the idea, the build, the launch, and the support afterwards.

The character of this era, and in practice it is one person changing hats all day. The bar for good enough went up at the same moment the headcount went down.

Doing exactly what you literally asked for and completely missing the point of it.

You asked for the button to be centred, and now nothing else on the page is. Technically correct. Practically useless.

Sending the unchanging chunk of your prompt over and over at a discount, instead of paying full price every time.

The biggest saving most solo builders never switch on. Put the parts that never change first and the parts that do change last, and the savings arrive without you rewriting a word.

An attack where something the AI reads — a web page, a file, a comment, a support ticket — contains instructions, and the AI just does what they say.

Harmless while the AI can only talk. Genuinely dangerous the moment it can act: send the email, delete the file, spend the money.

A mildly rude name for somebody who prompts all day and checks nothing.

The insult is not about skill level. It is about the missing second half of the job.

Looking things up in your own documents and handing the relevant bits to the AI at the moment you ask, instead of hoping it already knew.

From a 2020 research paper, now the normal way to make a general-purpose AI answer questions about your particular stuff. Most of the time it fails, it failed at the looking-up part, not the answering part.

Running an AI at the same job over and over until the answer stops changing.

Named for a pattern Geoffrey Huntley made popular. It works better than it has any right to, as long as the job has a finish line the machine can check. Without one it loops forever, confidently.

One round of handing the AI its own work back, along with what is wrong with it.

Rounds are cheap, and they are not equal. One and two usually fix how things look. The real structural mistake tends to surface on three or four, once the obvious problems stop hiding it.

A second sorting step that reorders search results by how well they actually answer the question.

Search wide, then sort hard. The top five straight out of a search are often worse than the top twenty sorted back down to five.

Research, Plan, Implement: the AI reads the code and writes a plan first, and is not allowed to change anything until you say go.

Credited to Dex Horthy. The whole discipline is refusing to let it start typing during the reading part, which it will always want to do.

Explaining your problem out loud, line by line, to something that cannot answer, until you spot the mistake yourself.

Made famous by The Pragmatic Programmer in 1999. The duck has since been promoted: it answers now, which helps about half the time.

A change that fixes the bug you reported and quietly breaks something nearby that was working fine.

The reason later rounds need the same care as the first. The AI improves the thing you pointed at and has no stake at all in the thing you did not.

An unasked-for, AI-written change that takes longer to read than the fix would have taken to write.

The bill for cheap generation. It does not disappear. It moves to whoever has to read the thing.

Registering a package name that AIs commonly invent, so that when one invents it, what arrives is the attacker's code.

Built on typosquatting, where the bait is a spelling mistake. Here the bait is the machine's imagination. It is the reason to check every unfamiliar add-on against the version you actually installed, before you read a line of it.

What builds up when the code keeps changing and the written plan does not.

Invisible until it is expensive. The plan stops describing the thing, and everybody keeps reading it anyway.

Writing the plan down properly first and treating it as the real thing, with the code as something generated from it and checked against it.

Usually shortened to SDD and done in four steps: write the spec, plan it, break it into tasks, build it, with a person checking at each step. It caught on because AI writes code well and guesses what you meant badly.

A second AI sent off to do one small job on its own, which comes back with only the answer.

The main defence against context explosion. The main one keeps the plan; the little one keeps the mess.

The things you can actually watch happen that tell you the job is finished.

Not the same as describing the feature. “Customer A gets zero rows belonging to customer B, and here is the test that proves it” is a success criterion. “Make it secure” is a wish.

Standing instructions that apply to every message in a conversation, sitting behind whatever you type.

Where the role, the rules and the hard nos live. Every word of it gets paid for on every single message, which is a good reason to edit it like a contract.

A dial for how much the AI is allowed to be random when picking its next word.

Turn it down for code and for pulling facts out of documents. Turn it up for ideas. It does not make the machine more right or more wrong, only more or less predictable.

The unit an AI reads and writes: roughly a short word, or a piece of a longer one.

Everything is priced, capped and counted in tokens, which is why the working unit of these systems is not the sentence but the scrap.

An AI actually doing something — running a command, reading a file, sending a request — instead of only writing about it.

Also called function calling. It is the line between a machine that describes an action and one that takes it, and it is exactly the line where prompt injection stops being a theory.

When an AI builds most of a feature in an hour, and the last little bit takes days.

What is left is hard precisely because you did not build the first part and cannot find your way around it. The time you saved was borrowed, not earned.

Almost everybody uses these tools every day, and trust in what they hand back keeps falling.

In Stack Overflow's 2025 survey, use climbed to eighty-four percent while active distrust of the results went from thirty-one to forty-six percent in a single year. The gap between what people use and what people believe is the whole opportunity.

The closest thing this crowd has to a motto: use the tools hard, then check what comes back without mercy.

The short version is shorter still. It runs is not it works.

Somebody who builds working software mostly by telling AI what to do, and takes full responsibility for the result anyway.

The extra letter is doing real work. A coder writes the code; a coader conducts it. The craft is not typing, it is framing the job, setting the limits, and checking. You judge one by how little they have to throw away, not by how much they produce.

Building software by describing what you want to an AI and taking what it gives you, without reading much of it line by line.

Andrej Karpathy coined it in February 2025, in the spirit of giving in to the vibes and forgetting the code exists. He meant low-stakes weekend messing about. It has since been stretched over serious paid work, which is why about half of every argument about vibe coding is really an argument about the definition. Collins made it their 2025 Word of the Year. The same year, Merriam-Webster picked slop.

Shortcuts and mess piling up as fast as a machine can type.

Ships in an afternoon, charges interest for a year. Ordinary shortcuts get taken by a person who knew they were cutting a corner. This kind gets taken by nobody in particular.

A proposed name for doing all of this properly: the same tools, but somebody actually reads the code.

Coined mostly by people tired of explaining that yes, they do read it. Expect the casual meaning of vibe coding to stick around, and the serious meaning to keep shopping for a new name.

A label for a build, a prompt or a session that failed, published with the post-mortem still attached.

The most useful thing anybody in this field can publish, and the rarest. Everyone shows the finished thing. The failure is where the knowledge you can actually reuse lives.

A chain of one-more-thing-first jobs that carries you further and further from what you sat down to do.

The phrase comes out of MIT lab lore in the 1990s. AI is unusually good at it, because every detour looks reasonable on its own, and nothing in the loop ever asks whether you are still solving the original problem.