went badly

State an agent carries across turns or sessions: recent history plus durable notes.

Distinct from the context window, which is what fits right now. Memory is what gets chosen to come back.

Development in which an AI agent runs a loop of gathering context, acting with tools, and verifying results, rather than answering one prompt at a time.

The grown-up sibling of vibe coding. The distinguishing feature is the verify step: an agent that cannot check its own work is just a very fast typist.

Generated output that is voluminous, generic, and low in value.

Merriam-Webster made slop its 2025 Word of the Year, the same year Collins chose vibe coding, which is a fairly complete summary of how that year went. The term now covers code, pull requests, documentation, and issues.

The set of files, systems, and behaviors a change can affect.

State it in the prompt and the agent stays in its lane. Leave it implicit and you get a one-line fix wrapped in four hundred lines of unrequested tidying.

An existing codebase with behavior that must be preserved while it is changed.

The real market and the hard problem. Most software work is brownfield, and most vibe-coding content pretends otherwise.

Copying the visible form of a practice without the understanding that makes it work.

From Richard Feynman's 1974 Caltech address on cargo cult science. The modern version is a repository full of configuration files copied from a blog post, none of which anyone can explain.

A model working through intermediate reasoning steps before producing an answer.

It improves accuracy on multi-step problems. It is not a transcript of the model's actual internals, and reading it as one is a category error.

A test written to capture what existing code actually does, rather than what it should do.

The essential tool for brownfield work. Before you let an agent near legacy code, you pin its current behavior, because otherwise you cannot tell a fix from a regression.

Output that is fluent, plausible, well-formatted, and incorrect.

The signature failure mode of the whole field. Human writing signals uncertainty through hedging and mess; generated writing does not, so fluency stops being evidence of anything.

The section of a prompt that bounds the solution space: stack, forbidden actions, invariants, and things the model must not touch.

The highest-leverage part of any prompt and the part most often skipped. A model will not infer your security boundary; if isolation is not in the constraints, it does not exist in the output.

The practice of deliberately assembling what enters a model's context window: retrieval, memory, system instructions, tool definitions, and prior turns.

It displaced prompt engineering as the term of art, and the shift matters. The leverage moved from how you phrase the request to what you put in front of the model, and just as importantly, what you keep out.

The degradation in output quality that occurs when an agent reasons over too much material at once.

More context is not more understanding. Past a point the signal you needed is buried in the material you added, and the usual defense is a subagent with a narrower job.

The decay of earlier decisions in a long session, as they fall out of the model's effective attention.

Also called context decay. The tell is an agent that silently contradicts a constraint you set forty turns ago, with no sign it ever knew.

The total number of tokens a model can attend to at once.

Bigger is not free. Cost, latency, and the difficulty of finding the relevant part all rise with it, which is why the skill is selection rather than accumulation.

The point at which a prototype that demonstrated beautifully cannot be extended into a product.

Everything before the cliff was generated; nothing before the cliff was designed. The fall is usually discovered when the second feature has to share state with the first.

Easy Approach to Requirements Syntax: five sentence patterns that turn vague requirements into testable, machine-parseable statements.

Created by Alistair Mavin and colleagues at Rolls-Royce and first published at the IEEE requirements engineering conference in 2009, long before anyone was prompting anything. The patterns are ubiquitous, event-driven, state-driven, optional-feature, and unwanted-behaviour. It is the cheapest upgrade available to most vibe coaders.

A numeric vector representing the meaning of a piece of text, such that similar meanings sit near each other.

The machinery under semantic search. Similar is not the same as relevant, which is why a second ranking pass usually earns its keep.

Writing the evaluation before the feature, so you can tell whether a change improved anything.

Test-driven development adapted for systems that do not produce the same answer twice. Without it, "the new prompt seems better" is the only quality signal you have, and it is not one.

Repeatable tests that score model or agent output against expected behavior.

Your regression suite for a system that will not give the same answer twice. The maturity move is replacing "that looked right" with a number you can watch over time.

Further training of an existing model on specific examples to specialize its behavior.

Usually the wrong first move. Most problems people bring to fine-tuning turn out to be retrieval or prompting problems that were cheaper to fix.

A project with no existing code, where the agent can generate freely.

Where the demos come from. Greenfield is largely solved, which is why greenfield demos impress nobody who has shipped anything.

A confident fabrication: a function, citation, parameter, or fact that does not exist.

The word is doing PR work for the machine, implying a perceptual error rather than a generative one. The model is not seeing things; it is producing plausible text, and sometimes plausible text is false.

Everything wrapped around a base model to make it an agent: system prompts, tool wiring, control flow, retries, and memory glue.

Also called scaffolding. Increasingly, gains on long-horizon tasks come from a better harness rather than a smarter model, which is why practitioners talk about the harness more than they talk about the weights.

The named point in a pipeline at which a person is accountable for what ships.

If you cannot point to yours, you do not have one. The gate is a person and a moment, not a policy document.

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

"Add login" is wildly underspecified, and the defaults a model picks to fill the gap are rarely yours. Drift is not the model being wrong; it is the model being reasonable about something you never said.

The core cycle of vibe coding: generate, evaluate, refine, repeat, until output converges on intent.

The whole practice lives here. Everyone shows you pass one; almost nobody publishes pass four, which is where you find out whether you understood the problem.

Model Context Protocol: an open standard for connecting AI agents to external tools and data sources.

Introduced by Anthropic in late 2024 and now the de facto plug standard. If a tool does not speak it, that tool is effectively outside the ecosystem.

Assembling all context, constraints, and reference material before the first prompt.

Borrowed from professional kitchens, where it means everything is on the board before the burner goes on. The failure mode it prevents is identical in both trades: improvising mid-service because you did not prep.

Sending different work to different models: cheap and fast for boilerplate, frontier for architecture, specialized for security review.

The tech-lead move applied to machines. Once you route, you stop having a favorite model and start having a workflow, and the orchestration layer quietly becomes your real IDE.

A single builder covering product, architecture, implementation, deployment, and support at a scope that recently required a small team.

The defining archetype of the era, and usually a rotation through several working styles rather than a single one. The bar for shippable rose at the same time the headcount fell.

Producing output that satisfies the literal request while missing its purpose.

You asked for the button to be centered and now nothing else on the page is. Technically compliant, practically useless.

Reusing a fixed prefix across many model calls to cut latency and cost.

The single biggest cost lever most solo builders never switch on. Put the stable material first and the variable material last, and the savings arrive without changing a word of the prompt.

An attack in which untrusted content carries instructions that an agent then follows.

The content can be a web page, a file, a code comment, or a support ticket. Harmless when the agent can only talk; genuinely dangerous the moment it can act.

Mildly derisive term for someone who prompts constantly and verifies nothing.

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

Retrieval-augmented generation: pulling relevant documents into the context window at query time instead of relying on trained-in knowledge.

Introduced in a 2020 research paper and now the standard way to make a general model answer questions about your particular data. Most RAG failures are retrieval failures wearing a generation costume.

Running an agent repeatedly on the same task until the output converges.

Named for a pattern popularized by Geoffrey Huntley. It works better than it should, and it works best when the task has a checkable finish condition; without one, it loops forever, confidently.

One iteration of feeding a model its own output plus a correction.

Passes are cheap and non-linear. Passes one and two usually fix cosmetics; the structural error tends to surface on three or four, once the obvious problems stop hiding it.

A second scoring pass that reorders retrieved candidates by relevance to the actual question.

Retrieve broadly, rerank narrowly. Top-five by raw vector similarity is frequently worse than top-twenty reranked down to five.

Research, Plan, Implement: a three-phase agentic method in which the agent studies the codebase and writes a plan before it is allowed to edit anything.

Attributed to Dex Horthy. The discipline is in refusing to let the agent start typing during the research phase, which it will always want to do.

Explaining a problem aloud, line by line, to an inanimate object until the flaw becomes obvious.

Popularized by The Pragmatic Programmer in 1999. The duck has been promoted: it now answers, which is an improvement roughly half the time.

A change that fixes the reported bug and quietly breaks a working behavior nearby.

The reason later refinement passes need the same test discipline as the first. The model optimizes what you pointed at and has no stake in what you did not.

An unrequested, agent-authored pull request that costs a maintainer more to review than the fix would have cost to write.

The externality of cheap generation. The cost does not disappear; it moves to whoever has to read it.

The practice of registering a package name that AI models commonly hallucinate, so that the fabricated import resolves to attacker-controlled code.

Built from typosquatting, where the bait is a misspelling; here the bait is the model's imagination. It is the reason to verify every unfamiliar import against your installed version before reading a single line of the implementation.

The divergence that accumulates when code evolves and its specification does not.

Invisible until it is expensive. The spec stops describing the system, and everyone keeps consulting it anyway.

A method in which a versioned, structured specification is the source of truth and code is a generated artifact verified against it.

Usually shortened to SDD and run in four phases: specify, plan, tasks, implement, with a human checkpoint at each. It went mainstream because agents write code well and guess intent badly.

A scoped agent spawned for one narrow task, running its own context and reporting a result back.

The main structural defense against context explosion. The parent keeps the plan; the child keeps the mess.

The observable, checkable conditions that define a finished piece of work.

Distinct from a description of the feature. "A user from tenant A receives zero rows from tenant B, proven by a test" is a success criterion. "Make it secure" is a wish.

Standing instructions that apply to every turn of a conversation, separate from the user's message.

Where role, constraints, and refusals live. Everything in it is paid for on every single call, which is an argument for editing it like a lease.

A setting controlling how much randomness a model applies when choosing its next token.

Low temperature for code and extraction, higher for ideas. It does not make a model more or less correct, only more or less predictable.

The unit a model reads and writes: roughly a short word or word fragment.

Everything is priced, limited, and measured in tokens, which is why the practical unit of an AI system is not the sentence but the fragment.

A model invoking defined external functions rather than only producing text.

Also called function calling. It is the line between a model that describes an action and an agent that takes one, and it is the line at which prompt injection stops being theoretical.

The pattern in which an agent produces ninety percent of a feature in an hour and the last ten percent takes days.

The remaining work is hard precisely because you did not build the first ninety percent and therefore cannot navigate it. The time saved was borrowed, not earned.

The condition of near-universal daily use of AI coding tools alongside falling trust in their output.

In Stack Overflow's 2025 survey, adoption climbed to eighty-four percent while active distrust of accuracy rose 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 community's nearest thing to a motto: use the tools aggressively, then check the output without mercy.

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

A builder who produces working software primarily by directing AI agents rather than typing implementation code, and who accepts full ownership of the result anyway.

The extra letter is load-bearing. A coder writes the code; a coader conducts it. The craft is not syntax but framing, constraint, and verification, and the measure of one is not how much code they generate but how little they have to throw away.

Building software by describing intent to an AI model and accepting the generated code with little or no line-by-line review.

Coined by Andrej Karpathy in February 2025, in the spirit of giving in to the vibes and forgetting the code exists. It was a description of low-stakes weekend hacking. It has since been stretched to cover serious production work, which is why roughly half of every argument about vibe coding is actually a disagreement about the definition. Collins named it their 2025 Word of the Year; the same year, Merriam-Webster picked slop.

Technical debt accumulated at generation speed.

Ships in an afternoon and compounds for a year. Ordinary technical debt is incurred by a person who knew they were cutting a corner; vibe debt is incurred by nobody in particular.

Proposed successor term for the disciplined practice: the same tools, with actual review.

Coined largely by people tired of explaining that they do read the code. Expect the casual sense of vibe coding to survive and the professional sense to keep shopping for a new name.

Category label for a build, prompt, or session that failed, published with the diagnosis intact.

The most credible thing a practitioner can publish and the rarest. Everyone shows the finished build; the failure is where the transferable knowledge actually lives.

A chain of prerequisite tasks that carries you further and further from the thing you set out to do.

The term traces to MIT AI Lab lore of the 1990s. Agents are unusually good at yak shaving, because each detour looks locally reasonable and nothing in the loop asks whether you are still solving the original problem.