Glossary
Plain-language meanings of the terms this course uses. Each term links to the concept, tutorial or other terms that say more.
- Agentic task
Work you hand to an AI assistant as an outcome rather than a single question. It plans and takes the steps itself, such as searching, reading files or creating new ones, and many steps happen before you have read anything. Also called delegated multistep work.
See also: Concept: Answering versus acting: delegated multistep work (members’ Studio); Glossary: Bounded task; Glossary: Approval gate (approval point); Toolkit: Delegating multistep work: Claude Cowork and tasks (members’ Studio)
- Approval gate (approval point)
A point, chosen before the work starts, where the work must stop and wait for your yes. Put one before anything that would be hard to undo, reaches outside the task, or would pass the output on. An approval protects you only if you read what you approve.
See also: Concept: Answering versus acting: delegated multistep work (members’ Studio); Concept: Recurring work: when no one is watching (members’ Studio); Glossary: Delegation brief; Concept lab: Delegation approvals: open-ended, scoped, and least access (members’ Studio)
- Artifact
A piece of work the assistant produces as a separate, self-contained item, such as a document, a table or a diagram, that you can view, edit and keep apart from the conversation. It is still a draft until you have checked it.
See also: Toolkit: Artifacts: drafts you can see, edit and share (members’ Studio); Glossary: Verify
- Bounded task
A task with one goal, one output and a clear finish, written down before it starts, together with what it may use and what it may not touch. Narrow scope is what makes the review possible.
See also: Concept: Answering versus acting: delegated multistep work (members’ Studio); Glossary: Delegation brief; Glossary: Scope creep
- Challenge
Part of Verify. You ask for alternatives, counter-arguments, unsupported statements and what might be missing. It helps you find problems; it does not replace checking against your sources.
See also: Glossary: Verify; Glossary: Self-review; Concept: Fluent is not the same as true (members’ Studio)
- Chat history
The record of your past conversations kept in your account. What is kept, for how long, and whether an assistant can look back through it all depend on the tool and its settings. History is not the same as memory, and neither is context you can rely on until you can see where it came from.
See also: Glossary: Memory (saved memory); Concept: What the model has in front of it (members’ Studio); Toolkit: Memory, chat search and chat history (members’ Studio)
- Citation (reference)
A pointer from a statement to its source: a document, a section, a page or an exact quotation. In a model's reply, a citation is generated text like everything else. It can point to the right place, the wrong place or nowhere, so you open the source and find the sentence.
See also: Concept: Grounding: where each statement comes from (members’ Studio); Glossary: Provenance; Glossary: Hallucination (fabrication); Concept lab: Evidence audit: tracing claims to their sources (members’ Studio)
- Compounding error
An error or assumption made at an early stage of multi-stage work and carried forward. Each later stage builds on it, it looks more settled each time it is repeated, and caveats drop out along the way.
See also: Concept: Errors compound across stages (members’ Studio); Glossary: Self-review; Glossary: Verify
- Connector
A link that lets an AI assistant work inside another service, such as email, a calendar or a document store, on your behalf. It is a standing permission: it stays in place until you take it back, whether or not you are watching.
See also: Concept: Granting access: what you are really handing over (members’ Studio); Glossary: Permission scope; Glossary: Revocation; Toolkit: Connectors: what you let Claude reach, and how to take it back (members’ Studio)
- Context
Everything the model has in front of it when it writes a reply: your request, the conversation so far, anything attached or pasted, and any standing material such as workspace files, saved memory or standing instructions. It works from this and from its training, nothing else.
See also: Concept: What the model has in front of it (members’ Studio); Glossary: Context window; Concept: The request shapes the answer (members’ Studio)
- Context window
The limit on how much text a model can take into account at once for a single request, including the conversation so far and any attached material. Very long conversations or very large documents can go beyond it, and then not everything can be used.
See also: Glossary: Context; Concept: What the model has in front of it (members’ Studio)
- Delegation brief
The written instruction for an agentic task: the goal, the scope, what it may touch and what it may not, how to treat untrusted content, the approval points, and what it must list when it finishes.
See also: Glossary: Bounded task; Glossary: Approval gate (approval point); Glossary: Least privilege; Toolkit: Delegating multistep work: Claude Cowork and tasks (members’ Studio)
- Evidence matrix
A simple table for your own notes. For each statement you check, it records the reference the reply gave, the source you actually found, a label (Known, Inferred, Unknown or wrong) and what you will do about it: keep, reword, correct or remove.
See also: Glossary: Known / Inferred / Unknown; Glossary: Provenance; Concept lab: Evidence audit: tracing claims to their sources (members’ Studio)
- Frame
The first habit of the operator framework. Decide what you want before you ask, and write it down as an assignment in seven parts: objective, role, context, assumptions, boundaries, output format and review criteria.
See also: Operator framework: Frame, Ground, Verify, Govern (members’ Studio); Concept: The request shapes the answer (members’ Studio); Glossary: Ground
- Govern
The fourth habit of the operator framework. You stay accountable: nothing is used, shared or acted on until you have approved it. It includes Decide (make the call yourself and be able to explain it) and Document (record what worked so it can be repeated and checked).
See also: Operator framework: Frame, Ground, Verify, Govern (members’ Studio); Glossary: Verify; Glossary: Approval gate (approval point)
- Ground
The second habit of the operator framework. Give the assistant approved, relevant material to work from, tell it to stay within that material, and know where each statement is supposed to come from.
See also: Concept: Grounding: where each statement comes from (members’ Studio); Glossary: Frame; Glossary: Verify
- Grounding
Asking a model to work from material supplied with the request, such as documents you attach or pages a search found, rather than from general patterns, and being able to see which statements came from that material.
See also: Concept: Grounding: where each statement comes from (members’ Studio); Glossary: Provenance; Glossary: Citation (reference); Toolkit: Files, documents and research (members’ Studio)
- Hallucination (fabrication)
A statement that sounds right but has no support: a figure nobody measured, a source that does not exist, a quotation never said, or a detail about a client you never supplied. It happens because the model writes plausible text, and the risk is highest where it has to fill a gap.
See also: Concept: Fluent is not the same as true (members’ Studio); Glossary: Citation (reference); Glossary: Verify
- Knowledge cutoff
The date the training data stops. On its own, a model may not know about anything that happened after it, unless you supply the material or it has a tool that searches for it.
See also: Glossary: Training data; Concept: What a language model does (members’ Studio)
- Known / Inferred / Unknown
Three labels for any statement that matters. Known: a supplied source states it, and you have found the sentence. Inferred: a reasonable reading of what the sources state, labelled as an inference. Unknown: the sources do not settle it, so it is a question to ask, not a finding. If a source says the opposite, the statement is wrong: correct it.
See also: Concept: Grounding: where each statement comes from (members’ Studio); Glossary: Evidence matrix; Concept lab: Evidence audit: tracing claims to their sources (members’ Studio)
- Language model
A computer program trained on a very large amount of text. It writes a reply a small piece at a time, choosing a likely continuation from learned patterns and from what you have given it. It is not looking up a stored, checked answer.
See also: Concept: What a language model does (members’ Studio); Concept: Fluent is not the same as true (members’ Studio); Glossary: Training data
- Least privilege
Granting the narrowest access that does the job, for the shortest time: one practice folder rather than all your files, reading rather than changing, and an account that holds only what the work needs.
See also: Concept: Granting access: what you are really handing over (members’ Studio); Glossary: Permission scope; Concept lab: Delegation approvals: open-ended, scoped, and least access (members’ Studio)
- Memory (saved memory)
Notes an assistant keeps from earlier conversations and brings into new ones, where a tool offers it and it is switched on. It is a selection, not a transcript, and it can be incomplete, out of date or wrong. If something matters, supply it again or put it in the workspace.
See also: Concept: What the model has in front of it (members’ Studio); Glossary: Chat history; Toolkit: Memory, chat search and chat history (members’ Studio)
- Permission scope
What a connection is actually allowed to do. Permission to read lets it search, open and summarize. Permission to change lets it send, edit, create, share or delete. What you ask for and what you have permitted are different things: the permission decides what it could do.
See also: Concept: Granting access: what you are really handing over (members’ Studio); Glossary: Least privilege; Glossary: Connector
- Project (workspace)
A workspace that holds the background material and standing rules for one piece of work, so every conversation inside it starts with them. It makes replies easier to check, not correct.
See also: Glossary: Project knowledge; Concept: One assignment, a workspace, a procedure, a schedule (members’ Studio); Toolkit: Projects: a workspace with its own files and instructions (members’ Studio)
- Project knowledge
The reference files you add to a workspace so the assistant can draw on them in every conversation there. You supplied them, so you can check a reply against them. It can still misread a file, miss part of it or add things that are not in it.
See also: Glossary: Project (workspace); Concept: What the model has in front of it (members’ Studio); Toolkit: Projects: a workspace with its own files and instructions (members’ Studio)
- Prompt
The request you send: one assignment, written for one piece of work. A framed prompt says what the work is for, who it is for, what to avoid and how you will judge the result.
See also: Glossary: Frame; Concept: The request shapes the answer (members’ Studio); Concept: One assignment, a workspace, a procedure, a schedule (members’ Studio)
- Prompt injection
Instructions hidden in material the assistant reads, such as a web page, a document or an email, that it may follow as if they came from you. Treat everything it reads as material, never as orders, and limit what untrusted material can reach and what can act on it.
See also: Concept: Answering versus acting: delegated multistep work (members’ Studio); Concept: Recurring work: when no one is watching (members’ Studio); Glossary: Least privilege
- Provenance
The trail from a statement back to its source: which document, which section, which sentence. A statement with provenance can be checked in seconds. A statement without it can only be trusted or doubted.
See also: Concept: Grounding: where each statement comes from (members’ Studio); Glossary: Citation (reference); Glossary: Evidence matrix
- Revocation
Taking access back. Disconnecting stops future use from the assistant's side, but the other service may keep the permission until you remove it there too, and material already retrieved may remain in past conversations. Know how you would do it before you connect anything.
See also: Concept: Granting access: what you are really handing over (members’ Studio); Glossary: Connector; Toolkit: Connectors: what you let Claude reach, and how to take it back (members’ Studio)
- Scheduled task
A task that runs the same instructions again and again at a time you chose, without you there. Every run is new output from new inputs, and quality can drift, so the checks are designed in advance: an approval gate, a review plan and a stop date.
See also: Concept: Recurring work: when no one is watching (members’ Studio); Glossary: Stop date; Toolkit: Scheduled tasks: recurring work you set up once and review every time (members’ Studio)
- Scope creep
When delegated work does more than you asked, or something different, because the task left room. A bounded task and a clear list of what it may touch are the defence.
See also: Glossary: Bounded task; Concept: Answering versus acting: delegated multistep work (members’ Studio)
- Self-review
Asking the assistant to check its own output, or running the same material through it again. It can help you find problems, but it works from the same inputs and can share the same blind spots. It is not independent verification.
See also: Concept: Errors compound across stages (members’ Studio); Glossary: Challenge; Concept: Fluent is not the same as true (members’ Studio)
- Side effect
Anything delegated work changes beyond the output you wanted: files created, changed, moved or deleted, or something sent or submitted. Review what it touched, not only what it produced, and note "none" when there were none.
See also: Concept: Answering versus acting: delegated multistep work (members’ Studio); Glossary: Approval gate (approval point)
- Skill (written procedure)
A written procedure for a kind of task you do the same way again and again, kept so it can be used whenever that task comes up. It holds steps, not facts, and it is tested, not trusted: run it, check it, revise it, run it again.
See also: Concept: One assignment, a workspace, a procedure, a schedule (members’ Studio); Toolkit: Skills: a reusable procedure Claude can follow (members’ Studio)
- Standing instructions
Instructions that apply to more than one request, such as the rules for a workspace or instructions that apply across an account. They make a behavior more likely, not certain, so you test whether each one was honored.
See also: Concept: What the model has in front of it (members’ Studio); Glossary: Project (workspace); Toolkit: Projects: a workspace with its own files and instructions (members’ Studio); Example personal instructions (members’ Studio)
- Stop date
The date a recurring task ends unless you decide, in writing, to renew it. Together with a review plan and pause triggers, it is how recurring work stays under your control.
See also: Concept: Recurring work: when no one is watching (members’ Studio); Glossary: Scheduled task
- Training data
The very large body of text a model learned its patterns from before you ever used it. It gives general patterns up to a cutoff date, not checked facts about you, your client or your field.
See also: Concept: What a language model does (members’ Studio); Glossary: Knowledge cutoff; Concept: What the model has in front of it (members’ Studio)
- Verify
The third habit of the operator framework. Check the output against your sources. For each statement that matters: does it come from the material, is it a reasonable inference, or is it unsupported? It includes Challenge.
See also: Concept: Fluent is not the same as true (members’ Studio); Glossary: Known / Inferred / Unknown; Glossary: Challenge