THE CLASSROOM / FOUR CORE CLASSES
Each class has a learning outcome, a practical exercise and a piece of evidence to keep. The exercises use your chosen tools; you can also complete their planning steps without installing software.
01 / FOUNDATIONS
Learning outcome Explain the difference between a model output, a source-backed claim and a decision.
Write the intended user, input, useful output and consequence of an error. A brainstorming draft and a customer-facing decision need different checking. Decide what success means before you inspect a persuasive result.
An answer can sound specific while containing invented details. Open the cited source, check that it supports the exact claim and preserve its date and context. A citation-shaped string is not evidence by itself.
Create a small set of representative cases, including an incomplete input and an answerable question. Record omissions, unsupported statements and corrections. Keep a baseline so a changed prompt can be compared rather than merely admired.
YOUR PRACTICAL CLASS
Choose a low-stakes topic you know. Ask a chosen tool for a short explanation. Mark each factual claim, inspect its support and rewrite one uncertain sentence. Do not use private customer information.
A one-page evidence sheet: the question, original answer, checked sources, corrections and a short explanation of what remains uncertain.
What evidence would change your mind about whether the output is good enough?
02 / CREATIVE WORKFLOWS
Learning outcome Describe an image workflow as connected operations and make a controlled comparison.
A visual workflow links distinct operations: loading a model, encoding a prompt, sampling and decoding an image are common examples. Identify what each connection carries. A node graph makes dependencies visible; it does not make every output correct.
Record the tool and model versions, workflow file, seed where available and relevant settings. A seed helps compare runs within a controlled environment, but does not guarantee identical bytes across versions, hardware or implementations.
Start with a supported small workflow. Keep other settings fixed while changing one parameter or prompt element. Compare against your stated goal, including unwanted changes. Save both successful and failed examples so the next choice has a reason.
YOUR PRACTICAL CLASS
On your own supported AI image tool, create a neutral still-life baseline. Change one prompt detail or setting. Compare the two outputs using three criteria: composition, instruction following and visible artefacts.
A workflow diagram, a settings record and a two-column comparison with your next experiment. Use original or licensed input material.
Which parts of your result can be recreated, and which depend on an environment you have not recorded?
03 / LOCAL SYSTEMS
Learning outcome Prepare a resource and data-boundary plan for a small AI workload.
Model size, numerical precision, activations, image or context dimensions and concurrency can affect memory demand. A model loading successfully does not establish useful latency or sustained throughput. Measure the workload you actually intend to use.
Use the official documentation for the chosen release and hardware. Keep dependencies isolated where supported, record versions and assess custom-node or extension sources before running their code. A downloaded workflow may depend on executable components.
Local inference can keep selected processing near the data, but local location alone does not establish security. Define user permissions, retrieval access, external calls, logging and retention. When a model or network is unavailable, choose a documented fallback rather than silently sending data elsewhere.
YOUR PRACTICAL CLASS
Write a deployment worksheet without installing anything: intended workload, available resources, supported software route, allowed data, external dependencies and a failure response. Then plan one small measurement you would run.
A resource inventory, a data-flow sketch and a test plan with quality, time and resource observations. Avoid assuming another author's benchmark applies to your setup.
What would cause you to stop the pilot instead of increasing its scope?
04 / AGENTIC WORK
Learning outcome Design a bounded agent-assisted task with a tool contract and an evaluation plan.
An assistant can prepare text. An agentic workflow may choose steps and call tools. Writing a draft, reading an approved record and changing an external system have different consequences. Define which actions are allowed for this role.
Retrieved pages and documents can contain instructions that conflict with the user's task. Keep untrusted context separate, scope retrieval and enforce tool permissions in application code. A prompt that says “be safe” is not an access-control system.
Specify required inputs, the expected result, approved tools, limits and when a person takes over. Test missing data, an unavailable tool, conflicting sources and a malicious instruction in a document. Review quality, cost and handovers before expanding autonomy.
YOUR PRACTICAL CLASS
Map an enquiry-triage process using synthetic examples. Let the proposed agent classify a request and prepare a response, with no sending tool. Define a read-only tool's inputs, returned fields and errors. Add three cases that should trigger a handover.
A role description, tool contract, five-case evaluation set and a table showing which outputs need review before release.
Can the workflow fail safely while preserving useful context for the next person?
CHECK YOUR UNDERSTANDING
The self-check asks about the principles in these classes. Review the explanations, revisit a class and repeat the check when you can explain your answer.
Take the core self-check ↗