How models work
Context window
The context window is how much a model can take into account at one time: your instructions, the files you attach, the conversation so far and its own reply. What falls outside it the model does not see, and in a long conversation the earliest material can be summarized or dropped.
Embedding
An embedding is a list of numbers that stands for the meaning of a piece of text, so passages with similar meaning sit near each other. Embeddings let a search find the relevant passage by meaning rather than by exact words.
Fine-tuning
Fine-tuning is further training of an existing model on a smaller, specific set of examples to change how it behaves. It changes the model itself, where a prompt or a skill changes only the instructions it receives.
Generative AI
Generative AI is software that produces new text, images, audio or code in response to a request, by predicting what output fits the patterns it learned in training. It drafts; it does not look anything up unless it is given a tool that can.
Inference
Inference is a trained model at work: taking an input and producing an output. Training happens once, at the lab; inference happens every time anyone sends a message, and it is what usage-based pricing charges for.
Knowledge cutoff
The knowledge cutoff is the date after which a model's training data stops. Anything later, such as a rate change, a new rule or a new listing, the model does not know unless it searches or you supply it.
Large language model (LLM)
A large language model is a generative AI model trained on a very large body of text to predict the next piece of text, one token at a time. That one skill, at scale, is what lets it draft, summarize, translate and follow instructions; Claude and ChatGPT are assistants built on such models.
Multimodal
A multimodal model works with more than one kind of input or output: text, images, audio, video, or documents such as PDFs. It can read a photo of a sign-in sheet, describe a floor plan or answer a question about a scanned page.
Open-weight model
An open-weight model is one whose trained parameters are published, so anyone can download it and run it on their own hardware. Open weights are not open source: the training data and code can remain private.
Reasoning model
A reasoning model works through a problem in intermediate steps before it answers, instead of answering in one pass. It is slower and uses more tokens, and it tends to do better on multi-step math, planning and analysis.
Token
A token is the unit a language model reads and writes: a whole word, part of a word, a number or a punctuation mark. Length limits, speed and usage-based prices are counted in tokens rather than words.
Training data
Training data is the text, images and other material a model learned from before release. It shapes what the model knows, how it writes and what it gets wrong, and it does not change after training ends.
Working with an assistant
Artifact
An artifact is a standalone piece of output an assistant creates beside the conversation, such as a document, a table, a chart or a small working app, that you can view, edit, download or share. The name is Claude's; other assistants offer similar features under other names.
Context file
A context file is the one page an assistant reads first: who you are, what you do, the rules that never move, and where your files are. Ours is called
CLAUDE.md; the name does not matter, the discipline does.Custom instructions
Custom instructions are standing preferences saved in an assistant's settings, such as who you are, how you write and what to avoid, applied to every new conversation. Products give the feature different names.
Few-shot prompting
Few-shot prompting means including a few examples of the output you want in the prompt, so the model follows their pattern, tone and format. It is often the quickest way to get output that sounds like you.
Memory
Memory is an assistant feature that carries information from one conversation into later ones, either details you ask it to keep or details it picks up on its own. What it keeps, and how you review or delete it, depends on the product and your settings.
Project
A Project is a workspace inside the assistant that keeps a set of files and standing instructions together, so every conversation started inside it begins with the same context. In Claude the feature is called a Project.
Prompt
A prompt is what you send a model: the request, plus any instructions, examples and material that go with it. The same model does very different work for a vague prompt and a specific one.
Skill
A skill is a file, or a small folder of files, that tells an AI assistant how to do one kind of job: what to ask for, what order to work in, what the finished output looks like, and what it must never do. The assistant loads it when a task matches its description, so you do not have to re-explain the job each time.
System prompt
A system prompt is the set of instructions a model receives before the conversation begins, setting its role, rules and limits. Products write their own, and custom instructions, projects and skills let you add standing instructions of your own.
Temperature
Temperature is a setting that controls how much randomness a model uses in choosing each next word. Lower settings give more predictable wording, higher settings more varied wording, and most chat apps set it for you.
Agents and tools
AI agent
An AI agent is a model given a goal, tools and permission to take several steps toward the goal on its own, checking results and choosing what to do next. A chat answers a message; an agent works through a task.
API
An API, or application programming interface, is the way one piece of software asks another for data or action. AI vendors offer their models through APIs so other products can build on them, often under different terms and prices from their chat apps.
Computer use
Computer use is an agent's ability to operate a computer or browser the way a person does: reading the screen, moving the pointer, clicking and typing. It reaches software that has no connector, at some cost in speed and reliability.
Connector
A connector links the assistant to a tool you already use (mail, calendar, a drive, a CRM) so it can read from and, with your approval, write to that tool. Connectors run under your own account and permissions; the assistant sees what you let it see.
Harness
A harness is the software around a model that makes it a working agent: it supplies the tools, manages the context, enforces permissions and keeps the record of what was done. The same model can be careful in one harness and careless in another.
Human in the loop
Human in the loop means a person reviews or approves a step before an AI system's output takes effect. It is a decision about where the checkpoints go, not a general promise that someone is watching.
MCP
MCP, the Model Context Protocol, is the open standard that connectors are built on: a way for a tool to describe what it can do so any compatible assistant can use it. You do not need to configure it yourself to use a connector.
Tool use
Tool use, also called function calling, is a model's ability to call outside tools during a task: search the web, run a calculation, read a file, or query a calendar or a CRM. The model decides when to call a tool, and the result comes back into its context.
Workflow
A workflow is a fixed sequence of steps, some done by a model and some by people or other software, run in the same order each time. Unlike an agent, a workflow does not choose its own path.
Getting it right
AI detector
An AI detector is a tool that estimates whether a text was written by a model. Its results are probabilities, not proof: detectors flag human writing as machine-written and miss edited machine writing.
Benchmark
A benchmark is a standard public test used to compare models, such as a fixed set of questions in math, coding or reading. Labs report benchmark scores when they release a model.
Bias
Bias in AI is a systematic slant in a model's output, inherited from its training data or from how it was built, that treats people or groups unevenly. It can show up in word choice, assumptions or recommendations without any instruction to produce it.
Citation
A citation in an AI answer is a link or reference to where a claim came from. It shows where the model looked, not that the source says what the answer claims.
Eval
An eval, short for evaluation, is a repeatable test of an AI system's output against a set standard: sample tasks, the results you would accept and a way to score them. It tells you whether a change to a prompt, a skill or a model made the work better or worse.
Grounding
Grounding means tying a model's answer to specific source material, such as documents you supply or search results, rather than its general training. A grounded answer can be checked against the source it came from.
Hallucination
A hallucination is output a model presents as fact that is false or invented: a statistic, a quote, a case or a source that does not exist. It happens because the model writes what is plausible, and it reads as confidently as a correct answer.
Retrieval-augmented generation (RAG)
Retrieval-augmented generation is a method that searches a set of documents for passages relevant to a question and gives them to the model with the question, so the answer draws on them. It is how most "chat with your files" features work.
Sycophancy
Sycophancy is a model's tendency to agree with the user, praise the work or say what the user seems to want, even when that is wrong. Labs work to reduce it; it has not gone away.
Verification
Verification is checking an AI output against a source of truth before it is used: the number against the record, the quote against the document, the name against the file. A person or a separate check does it; asking the model whether it is sure is not verification.
Data, privacy and risk
AI disclosure
An AI disclosure is a statement telling a reader, viewer or client that AI was used to create or alter something. Some platforms, laws and brokerage policies require one, and the rules differ by medium and place.
Content credentials
Content credentials are tamper-evident metadata attached to an image, video or audio file recording where it came from and how it was edited, including whether AI was used. They follow an open standard from the Coalition for Content Provenance and Authenticity (C2PA).
Data retention
Data retention is how long a vendor keeps your conversations, files and outputs, and for what purposes. Deleting a chat from your screen does not always remove it from the vendor's systems at once; the retention terms say what happens and when.
Deepfake
A deepfake is audio, video or an image made or altered by AI to show a real person saying or doing something they did not. A cloned voice can be built from a short recording of the real one.
Guardrails
Guardrails are the rules and checks that keep an AI system inside set limits: what it may say, what it may do, what it must hand to a person and what it must refuse. They can be written into instructions, enforced by software around the model, or both.
Jailbreak
A jailbreak is a prompt written to get a model to ignore its safety rules or its maker's instructions. In a jailbreak the user attacks the model's rules; in prompt injection a third party attacks through the content the model reads.
Personally identifiable information (PII)
Personally identifiable information is any detail that identifies a person or can be tied to one: a name, an address, a phone number, an email, account and Social Security numbers, financial details. Privacy laws and contracts set rules for how it is collected, shared and kept.
Prompt injection
Prompt injection is an attack that hides instructions in content an AI system reads, such as a web page, an email or a document, so the system follows the attacker's instructions instead of yours. It is a leading risk for agents that browse, read mail or act in your accounts.
Shadow AI
Shadow AI is the use of AI tools at work outside the organization's knowledge or approved accounts: a personal subscription, a free browser extension, an app on a phone. The risk is less the tool than that no one knows what data went where.
Training on your data
Training on your data means a vendor uses what you type or upload to improve its future models. Whether it happens depends on the product, the plan and your settings, and personal and business plans often differ.
Virtual staging
Virtual staging is adding furniture, decor or finishes to a listing photo digitally, more and more often with AI. AI editing can also change what a photo shows about the property itself: a sky, a lawn, a wall, a view.
Zero data retention
Zero data retention is an arrangement, usually for business or API customers, under which a vendor does not store your inputs and outputs after processing them, apart from exceptions its terms name. It is typically a contract or account term, not a switch in a chat app.
The Practice's method
Client-owned delivery
Client-owned delivery means everything Stoa sells is installed in your own accounts and saved in your own folders: files you can open, edit, and keep. Stoa hosts none of what you buy, holds none of your business data, and cannot switch anything off.
Concept plate
A concept plate is a still image that carries one idea in the brand's own tokens: a diagram, a stat with its source, a short title card. It is rendered from code so every word on it is exact and checkable, and it carries no caption, no stock imagery, and no counterfeit interface.
Decisions lock
A decisions lock is the list of choices already made, written down before an audit begins, so the audit checks whether the work kept them rather than proposing new ones. Your typeface, your two colors, your margin, your venue.
Drift check
A drift check is the monthly question that keeps a written procedure or role true: "is this still how it is done?" Yes refreshes the review date; no reruns the capture against what changed.
Evidence lock
An evidence lock is the step where facts are captured and written down before anyone offers an opinion: the screenshots, the text sizes, the measured margins, the dates. Opinions come after, and each one cites the evidence it rests on.
/go path
A
/gopath is a short address on our own domain (runthepractice.com/go/<name>) that forwards to a destination with a label naming the surface that sent you: a class handout, a desk card, a story, a letter, a guide. The page you land on reads the label. It is printed as a QR plate or written as a link.Rung ladder
The rung ladder is the marketplace's price structure: Guide (free), Pack, Instrument, Category bundle, Library, Industry bundle, and The Complete Practice, with every price set by the rung rather than by the product. A bundle is always priced below the sum of its parts and above its largest part.
Runner mark
A runner mark says who does a step: PERSON, ASSISTANT, or EITHER. PERSON is for judgment, signatures, conversations, licensed acts, and anything regulated; ASSISTANT is for deterministic work from known inputs; EITHER is for whoever is at the desk.
Vintage
A number's vintage is its as-of date and its source, carried with the number everywhere it travels. A median price without a vintage is a rumor; with one, it is a fact an agent can repeat and a reader can check.
Educational, not professional advice. Nothing on this page is affiliated with or endorsed by its publisher.


