Glossary and cheat sheet
Every term in the course at three reading levels, the brief template and a prompt library.
You can look up any course term and copy a prompt when you need one.
You can look up any course term at your reading level and copy the brief template or a prompt from the library.
You have the course vocabulary at three levels of precision, the brief template and a library of prompts to adapt.
Glossary
- Model
A program that learned from lots of text and writes answers by guessing what comes next.
The trained engine inside an AI product. It predicts text from patterns it learned.
A neural network, here a transformer, whose weights define a next-token probability distribution.
- Token
A small piece of text, often part of a word.
The unit a model reads and writes. Limits are counted in tokens.
An element of the learned subword vocabulary. The unit of context length and of billing.
- Training data
The text the model learned from before you used it.
The material the model was built from. It stops at a fixed date.
The pre-training corpus plus post-training data. Fixed when the model is released.
- Knowledge cutoff
The date when the model's learning stopped.
The date after which the model knows nothing unless a tool or a file gives it the facts.
The latest date represented in training data. Later facts must enter through the context.
- Reasoning effort
How long the model thinks before it answers.
A Codex setting that trades time and allowance for more careful work.
A control over the budget of intermediate reasoning tokens generated before the answer.
- Context
Everything the model can see when it writes one answer.
Everything the model can see for one answer.
The full token sequence supplied as input for one generation: instructions, history, files read and tool results.
- Context window
The most the model can see at one time.
The maximum size of the context.
The maximum number of tokens, input and output together, that the model can attend to in one call.
- Compaction
Making old parts of a long chat shorter so the chat still fits.
Summarising older parts of a thread to fit the window.
Replacing earlier turns with a model-written summary to make room in the window, at the cost of detail.
- Prompt
The message you send.
The message you send.
The user-authored part of the input. It sits in the context beside system instructions and history.
- Tool
An action the harness can do for the model, such as reading a file.
An action the harness can run for the model.
A function the harness exposes to the model, described by a schema, that it can request to be run.
- Tool call
One request from the model to use a tool.
One request to run a tool.
A structured output naming a tool and its arguments. The harness runs it and appends the result to the context.
- Shell
A window where you type commands to the computer.
The command line.
A command interpreter. Codex uses it as a general-purpose tool to run programs and read or write files.
- PowerShell
The Windows window for typing commands. Codex uses it.
The Windows command line Codex uses.
The Windows shell and scripting language in which the Codex agent runs commands natively under the Windows sandbox.
- Harness
The program around the model. It gives the model its tools and keeps it safe.
The software around a model that builds context, runs tools and enforces permissions.
The runtime that assembles the context, exposes tools, executes tool calls, enforces the sandbox and permission mode, and keeps history.
- Agent
A model with tools that keeps working toward a goal.
A model, a harness and tools working in a loop toward a goal.
A system in which a model chooses actions through a harness and observes the results, over repeated steps, to reach a goal.
- Agent loop
Plan, act, look at the result, and repeat.
Plan, act, observe, repeat.
The cycle of reasoning, tool use and observation, repeated until a stop criterion is met, as in ReAct (Yao et al., 2022).
- Sandbox
A fence around Codex. It limits which files and networks it can reach.
The boundary that limits which files and network a command can reach.
An operating-system-enforced boundary on file and network access for commands the agent runs.
- Permission mode
A setting that says when Codex must ask you first.
The Codex setting that decides what needs your approval.
A preset pairing of sandbox mode and approval policy: Ask for approval, Approve for me, Full access or Custom.
- Approval
Your yes or no to one thing Codex wants to do.
Your yes or no to one action Codex wants to take.
A human decision on a single proposed action, raised when the action falls outside the sandbox policy.
- Local
Running on your own laptop.
Tools run on your laptop.
The harness and its tools execute on your machine. The model still runs on remote servers.
- Cloud
Running on OpenAI's computers.
Tools run on OpenAI's computers.
The harness and its tools execute in a remote environment run by the provider.
- Workspace
Your organisation's space in ChatGPT.
Your organisation's ChatGPT account space.
The organisation-level ChatGPT tenant. Admin policy and data terms attach to it.
- Project folder
The folder you open in Codex. Codex works inside it.
The folder you open in Codex; the sandbox centres on it.
The folder you open in Codex. The sandbox centres on it, and Codex searches it for AGENTS.md.
- Brief
A short note with six parts that tells an agent what to do.
A six-part task description.
A task specification: goal, inputs, output, constraints, termination condition and verification method.
- AGENTS.md
A file of rules that Codex reads at the start of each task.
A file of standing instructions Codex reads at the start of each task.
A persistent instruction file loaded into the context at run start, with the nearest file taking precedence and a 32 KiB combined limit.
- Skill
A recipe you save once and use many times.
A reusable recipe in a folder with SKILL.md.
A packaged procedure: SKILL.md with name and description, plus optional scripts, references and assets, loaded on demand.
- Plugin
A bundle of skills and links to other tools.
A bundle of skills and connections to other systems.
A distributable bundle of skills, MCP servers and hooks.
- MCP (Model Context Protocol)
A shared standard for plugging an agent into other systems.
A standard way to connect an agent to other systems.
An open client-server protocol over JSON-RPC by which a server exposes tools, resources and prompts to an agent.
- Verification
Checking an answer against a source or a rule.
Checking an output against a source or a rule.
Testing an output against an independent source, rule or executable check, rather than against the generator's own claim.
- Prompt injection
Hidden orders in a file or web page that try to steer Codex.
Text in a file or web page written to make an agent do something you did not ask.
Adversarial instructions put in content the agent reads, which the model may treat as if the user wrote them.
- Hallucination
A wrong answer given in a sure voice.
A confident error.
A fluent output that is not supported by the context or by fact, produced with no signal of low confidence.
- BPMN (Business Process Model and Notation)
A standard way to draw how a job moves from step to step.
A standard notation for drawing a business process as lanes, tasks, gateways and events.
An OMG standard, BPMN 2.0, for process models with an XML interchange format that tools such as viewers can render.
- P50
The amount of power a farm should beat half the time.
The generation level expected to be exceeded half the time.
The median of the forecast distribution of generation. Half of outcomes should exceed it.
- PPA (power purchase agreement)
A contract to buy power from a farm at agreed terms.
A long-term contract in which a buyer agrees to purchase power from a generator on set terms.
A bilateral contract for the sale of electricity, whose term sheet sets dates, prices and conditions that affect revenue.
- Three-way match
Checking that the invoice, the order and the delivery all agree.
Checking an invoice against its purchase order and receipt. In the lab, invoice against purchase order.
A payables control that matches invoice, purchase order and goods receipt before payment. The lab applies the two-way form.
- Container
A small sealed box that runs one program the same way on any computer.
A packaged program with what it needs to run, isolated from the rest of the computer.
An isolated process environment created from an image, sharing the host kernel.
- Docker
A tool that builds and runs containers.
The tool that builds and runs containers on your laptop.
A container platform. Docker Desktop on Windows can use the WSL2 backend.
The brief template
Use these six lines each time you give Codex a job. Fill in each line with a short answer.
Use these six lines each time you give Codex a task. A short answer for each line is enough.
Use these six lines as a specification checklist. Keep the labels so that omissions are visible.
Goal: ... Inputs: ... Output: ... Constraints: ... Done when: ... How to check: ...
Prompt library
Copy a prompt, then change the file names to match your task. Use Ask for approval and read each card.
Copy a prompt and adjust the file names. The first three work for any task. The others match the labs, and each tells Codex not to change the source files.
Adapt these prompts. Each states its constraints and its output path. The lab prompts assume the starter pack layout.
Show me a numbered plan for this task. List your assumptions and any missing information. Do not change any files until I approve the plan.
Check your work against the Done when line. Show the evidence for each key figure, with the file and the page or line. List anything you could not confirm.
Explain business-analyst/data-classification-policy.md in plain words for a new starter. Use five bullet points. Do not change any files.
Read business-analyst/supplier-onboarding-process.docx. Write BPMN 2.0 XML to outputs/supplier-onboarding.bpmn with one lane per role. Also write outputs/view-process.html, which shows the BPMN with the bpmn-js viewer loaded from a CDN. List every task and gateway, and anything in the text that is unclear. Do not change the source file.
Classify each field in business-analyst/customer-fields.csv using business-analyst/data-classification-policy.md. Write outputs/field-classification.csv with the columns field, level, reason, personal_information and needs_review. Mark needs_review for any field the policy does not settle. Do not change the source files.
Compare asset-manager/generation-2026-09.csv with asset-manager/budget-p50-2026-09.csv. Write outputs/variance.py, run it, and save outputs/variance.csv and a chart. Write outputs/summary.md. State each cause as a hypothesis and show the data that supports it. If matplotlib is missing, make an SVG chart with no extra packages. Ask before you install anything. Do not change the source files.
Read asset-manager/ppa-*.pdf. Write outputs/key-dates.csv with the columns document, item, date, page and quote. Write outputs/key-dates.ics for the same dates. Show how you worked out any relative deadline. Do not change the source files.
Read procurement/rfq-laptops.md and procurement/quotes/*.pdf. Score each quote with the method in the RFQ and show the method. Write outputs/quote-scores.csv and outputs/recommendation.md. Do not change the source files.
Read procurement/invoices/*.pdf and procurement/purchase-orders.csv. Write outputs/invoices.csv with one row per invoice. Match each invoice to its purchase order and check the GST. Write the exceptions to outputs/exceptions.md. Do not change the source files.
Read presentation/steering-committee-notes.md. Write outputs/steering-committee.pptx with six slides. Put detail in the speaker notes and use at most five bullets per slide. Use a Presentations plugin if one is available. Otherwise write a python-pptx script, and ask before you install anything. Do not change the source file.
Write a Dockerfile that uses nginx:alpine to serve docker/site/index.html. Write a compose file that maps port 8080. Put both files in outputs/docker/. Ask before you run docker and before you pull the image. Do not change the source files in docker/site/. After I check http://localhost:8080, stop and remove the container.
Check yourself
Which term names the boundary that limits which files and network a command can reach?
The sandbox.
The sandbox.
The sandbox, enforced by the operating system on commands the agent runs.
Which prompt in the library do you use when you do not trust a result yet?
The check your work prompt. It asks Codex to show its proof.
General: check your work. It asks for evidence tied to the Done when line.
General: check your work. It asks for evidence against the acceptance criteria and a list of unconfirmed items.
OtterBlock is not affiliated with OpenAI. Codex and ChatGPT are OpenAI products.