Glossary

Plain meanings for course words. In lessons, hover a dotted link for a short tip, or click through for the full note.

agent

Also called: agents, agent loop

An LLM that can also act by opening tools.

An agent is a helper that can call tools, read results, and continue. You steer a loop: ask, act, observe, check. Simple chat only writes text.

Related modules: Module 1: Mental model, Module 5: Tools & MCP

agent harness

Also called: harness, agent runtime

The workbench around an agent: rules, tools, files, memory, approvals, and checks.

An agent harness is the workbench around the model. It supplies rules, tools, files, memory, approval gates, logs, and checks. A chat app or coding agent may provide this workbench for you.

agentic framework

Also called: agent framework, agent frameworks

A builder kit for making a custom agent workflow.

An agentic framework is code builders use to join model calls, tools, state, and approval steps. Most everyday workflows do not need one. Choose one only when a ready-made harness cannot run the workflow safely.

allow list

Also called: allow-list, enabled tools

The short list of apps you turn on for a job.

An allow list is the set of tools you enable. Smaller lists reduce wrong picks and risky actions. Pair it with a never-list for send, delete, pay, and production changes.

Related modules: Module 5: Tools & MCP

approval

Also called: approvals, human approval

A human gate before a risky action ships.

Approval means a person reviews a plan, diff, or send before it happens. Use approvals for delete, pay, public posts, and production changes.

binder

Also called: binder lookup

Our plain name for files and sources you look up on demand.

The binder is where project facts live: BRIEF.md, SOURCES.md, wikis, and docs. Retrieval is looking something up in that binder when you need it.

Related modules: Module 6: Retrieval & grounding

BRIEF.md

Also called: project brief

A one-page summary of the problem for later reuse.

BRIEF.md is a short project summary many teams keep in a repo or wiki. It states the goal, limits, and key facts with citations. Attach it later instead of pasting a whole research crawl.

capstone

Also called: capstone lab

The final practice run of the full loop on real work.

The capstone is where you configure the loop, run one real workflow, verify, and keep what worked as a team default.

Related modules: Module 12: Capstone lab

citation

Also called: citations

A pointer to the source that supports a claim.

A citation links a claim to a page, file, or note. In this course, put links and dates in SOURCES.md and keep BRIEF.md claims tied to those sources.

compaction

Also called: handoff note, handoff notes

A short note so you can clear the desk and keep going.

Compaction shrinks a long chat into the facts you still need: goal, done checks, locked choices, sources, and open questions. Then you can restart clean without losing the thread.

Related modules: Module 7: Conversation & compaction

context loop

Also called: the loop

The course map: goal, desk, tools, results, memory, and check.

The context loop is the path work takes with an AI helper. Your goal and rules load onto the desk. The helper may use tools. Results return. You verify. Memory and notes feed the next turn.

Related modules: Module 1: Mental model

context pack

Also called: research pack

A short set of files you reuse: brief, sources, and limits.

A context pack is the take-home from deep research. At minimum keep a one-page summary and a sources list. Later modules attach the pack instead of re-chatting folklore.

Related modules: Module 2: Deep research

context window

Also called: context windows

The helper's desk for one session — only so much fits at once.

The context window is everything the helper can see in one session. Think of it as desk space for one homework session. Messages, rules, files, and tool results all share that desk. When it fills up, quality often drops before the app warns you.

deep research

Also called: fact-finding

Fact-finding before you start a project.

Deep research collects real pages and notes first. Then you build. Save a short pack with links and dates so later chats do not guess.

Related modules: Module 2: Deep research

delegation

Giving one clear job to a specialist helper.

Delegation splits a mega-task into smaller jobs with contracts: goal, inputs, tools, out of scope, done checks, and who merges. Parallel work needs safe file ownership.

desk space

A plain name for the context window — room on the helper's desk.

Desk space is our course name for the context window. Only so much fits at once. Keep standing rules lean. Pull in files only when needed.

Related modules: Module 1: Mental model

done checks

Also called: done checklist, acceptance criteria

A short list that proves the job is finished.

Done checks are testable signs of finished work. Write them before you generate. Use them when you verify. Vague goals make weak checks.

Firecrawl

One example research tool that can collect clean page text.

Firecrawl is one brand of research helper. Other browse or deep-research tools can do similar work. The habit matters more than the brand: save sources with dates into your pack.

Related modules: Module 2: Deep research

grounding

Tying answers to real sources you can check.

Grounding means the answer rests on evidence you provided or looked up. Ask for citations. Check numbers and dates against the source.

Related modules: Module 6: Retrieval & grounding

hallucination

Also called: made-up facts, invented facts

When the helper invents details that are not in your sources.

Hallucination is a plain risk: confident wrong facts. Reduce it with deep research, retrieval, citations, and verify. Never paste guesses into wall rules.

human craft

Your skill at writing clear asks, modes, and gates.

Human craft is the part only you own: specs, reverse prompting, approval choices, and judgment. Tools speed work. You aim it.

Related modules: Module 10: Human craft

LLM

Also called: large language model, language model

A text engine that reads words and writes words.

An LLM (large language model) predicts helpful next words from the text on its desk. It is not a database. It needs clear goals, rules, and sources to do real work well.

Related modules: Module 1: Mental model

MCP

Also called: Model Context Protocol

A shared way to plug apps into helpers — like a shared app menu.

MCP stands for Model Context Protocol. It is a shared shape for listing tools so different helpers can use the same connectors. You still choose which apps to turn on.

Related modules: Module 5: Tools & MCP

memory

Also called: sticky notes

Sticky notes for next time — facts that should return later.

Memory stores facts across sessions. Keep durable truths. Prune secrets, one-day guesses, and huge tables. Prefer binder notes you control for project truth.

Related modules: Module 8: Memory systems

merge owner

Also called: merge-owner

The one person who combines specialist results.

When you delegate, name a merge owner. That person integrates outputs and resolves conflicts. Without a merge owner, parallel helpers collide.

never-list

Also called: never list

Actions the helper must not take without a person.

A never-list names banned or gated actions: send email, delete, pay, deploy to production. Put it in wall rules or the tool allow list setup.

observation

Also called: observations, tool result, tool results

What a tool sends back onto the desk after it runs.

An observation is a tool or search result added to the chat. It uses desk space. Huge dumps can bury your goal — summarize before the next turn.

Related modules: Module 1: Mental model, Module 5: Tools & MCP

playbook

Also called: playbooks, recipe card, recipe cards

A recipe card you pull out for one kind of job.

A playbook is a short guide for one workflow. It has a trigger, steps, and done checks. Put always-true rules on the wall. Put one-job steps on a recipe card.

Related modules: Module 4: Standing playbooks

prompt

Also called: prompts

The words you type to steer the helper this turn.

A prompt is your message. Strong prompts state the job, constraints, and sources. They do not replace wall rules, playbooks, or research packs.

Related modules: Module 10: Human craft

prune

Also called: pruning

Delete stale or unsafe notes from memory and rules.

Prune means clean house. Remove secrets, expired facts, and conflicting wall lines. Fresh desks and honest sticky notes beat a pile of junk.

RAG

Also called: retrieval augmented generation

A setup that looks up documents, then answers from them.

RAG means the system retrieves snippets, then the model writes with those snippets on the desk. It is retrieval plus generation. You still verify citations.

Related modules: Module 6: Retrieval & grounding

retrieval

Looking something up in a binder before you answer.

Retrieval means pulling the right page or note into the desk when you need facts. Attach or search your sources. Do not ask the helper to invent what your binder already holds.

Related modules: Module 6: Retrieval & grounding

reverse prompting

Also called: reverse prompt

Ask the helper what it needs before it starts work.

Reverse prompting surfaces missing facts and fuzzy goals early. The helper asks clarifying questions. You answer. Then it drafts. This cuts rework.

Related modules: Module 10: Human craft

shared workspace

Also called: team workspace, shared team workspace

The team project room for shared rules, sources, playbooks, tools, and ownership.

A shared workspace holds the files and defaults a team uses together. It names canonical sources, shared rules, tool setup, handoff formats, and who owns each output. Personal scratch chats stay outside it.

SOURCES.md

Also called: sources list

A list of links, dates, and why each source matters.

SOURCES.md holds evidence: URL, date accessed, and one line on why it matters. Refresh it when facts change. Use it when answers must be grounded.

standing context

Rules and files loaded before you type your ask.

Standing context is the quiet stuff already on the desk: wall rules, project notes, and sometimes memory. It shapes every answer until you change it.

standing instructions

Also called: wall rules, classroom wall rules

Rules on the classroom wall — loaded before you ask.

Standing instructions guide every answer in a space. They cover tone, limits, and what to do when something is unclear. They are not facts. Point to your research pack for proof.

Related modules: Module 3: System instructions

subagent

Also called: subagents, specialist

A specialist helper with its own desk and a clear job.

A subagent is like a specialist classmate. You write a brief. They work on their own desk. They return results. You merge. Do not let two helpers edit the same file without a merge owner.

system instructions

Also called: custom instructions, project instructions

Product name for standing instructions / wall rules.

Many apps call wall rules “system instructions,” “custom instructions,” or “project instructions.” Same job: always-on behavior, not a binder of facts.

Related modules: Module 3: System instructions

token

Also called: tokens

A small chunk of text the model counts as desk usage.

Tools often measure desk space in tokens. You do not need exact math. Know the idea: longer files and tool dumps cost more room on the desk.

tool

Also called: tools

An app the helper is allowed to open, like browse or read a file.

Tools let the helper act, not only chat. Each tool has a name and a short description. Keep the list small so the helper picks the right app.

Related modules: Module 5: Tools & MCP

user intent

Also called: intent, your goal

What you want done, clear enough to check.

User intent is your ask for this turn or job. Clear intent names the outcome, audience, limits, and done checks. Vague intent wastes desk space.

Related modules: Module 1: Mental model, Module 10: Human craft

verify

Also called: verification

Checking your work before you turn it in.

Verify means you check the output against done criteria before you trust it. Look at numbers, names, and sources. Fix, approve, or start over.

Search the course

Glossary