PRACTICAL AI / REVIEWED 2026-10-23
AI foundations: six useful lessons
Understand the tools, give a clear brief, and learn to check the result.
Define the task → Supply context → Try it → Check it → Improve it
Read and practice at your own pace. These activities use prepared examples; no AI service is connected. Completion records reflect your participation, not a professional qualification. This revised edition starts a new progress record.
01Know what you are usingChoose capabilities that fit the task.Open lesson
A model generates or interprets information. An app gives it an interface, instructions and access to features. An agent combines a model with tools to carry out a sequence of actions. The same model can behave differently in different apps.
Current systems can reason through problems, write code and work with supported images, audio or other files. Some can search, calculate or operate software. Availability depends on the product and its settings. Capability is not a guarantee of correctness.
For a current fact, look for a search-enabled tool and inspect its sources. For a calculation, use a calculator or executable analysis and check the inputs. For a document question, supply the document and ask for the supporting passage.
Try it
Choose one task you do regularly. Name the inputs, the capability it needs, and one way you could check the result.
02Give a useful briefMake the result easy to judge.Open lesson
Start with the task, the relevant material and the result you want. Add constraints that actually matter. An audience, example or tone can help; a long persona and a list of buzzwords are not prerequisites.
Separate instructions from the material being analyzed. Say what to do when information is missing. A short prompt can be excellent if it gives enough direction for that task.
Ask for a useful explanation, assumptions, evidence or a calculation when you need to review them. Reasoning models do not need a universal instruction to reveal every internal thinking step.
Try it
Rewrite “make this better” for one real task. Specify the change you want and how you will recognize a useful result.
03Keep context usefulDistinguish context, memory and sources.Open lesson
Tokens are small units of text, such as word parts or punctuation; exact boundaries depend on the model. The context window is the material available for the current response. Limits vary by model and app; some documented offerings support a million tokens or more. A larger window still does not guarantee complete recall.
Training knowledge comes from model development. Saved memory is an app feature that may retain selected information. Retrieval brings relevant material from files or a knowledge collection into the current request. These are different mechanisms.
Long conversations may be summarized, shortened or stopped at a limit. Keep an explicit working brief with current decisions, source names and open questions. If a detail matters, include it or ask the system to retrieve the passage instead of assuming it remains available.
Try it
Create a five-line handoff note for a task: goal, current decision, source, unresolved question and next action.
04Check the result against evidenceFind errors a fluent answer can hide.Open lesson
Separate claims, calculations and recommendations. Open cited sources and check whether they support the actual claim, including dates, units and scope. A citation can exist and still be irrelevant.
Use representative examples to evaluate quality. Include a missing-information case and an awkward case, not just an easy success. Agreement between two models is not independent proof if both repeat the same error.
Ask the system to flag uncertainty and explain assumptions, then verify the parts that matter. If evidence is unavailable, leave the point unresolved rather than filling the gap with a plausible story.
Try it
Take a short output and mark one fact to verify, one assumption to question and one recommendation requiring judgment.
05Use tools with clear boundariesDecide what an assistant may act on.Open lesson
Tools can let an assistant search, read files, run calculations, update records or send messages. Give it the access needed for the task, and define which actions need review before execution.
A web page or retrieved document may contain instructions aimed at the assistant. Treat that material as evidence to inspect, not authority to change the task or disclose information. This is the practical problem behind prompt injection.
For an early workflow, start with read-only access and drafts. Expand permissions only after testing the behavior, failure handling and ability to recover. Use approved environments for sensitive work.
Try it
For a meeting assistant, list what it may read, what it may draft and what it must ask before sending or changing.
06Make one workflow repeatableLeave with a small, testable plan.Open lesson
Choose a task you understand well enough to judge. Record the current effort and quality, then define a useful outcome. Use approved or fictional examples for a first test.
Try a small set of representative cases. Record missing facts, errors, review effort and whether the result was actually useful. Time savings are something to measure, not a promised percentage.
Keep what works, revise the brief and test again. Name the person responsible for reviewing the result and deciding whether to expand, revise or stop the workflow.
Try it
Write your task, success condition, test cases, reviewer and next decision in the brief builder below.
LEAVE WITH SOMETHING USEFUL
Your brief.
Your review plan.
Choose a starting point or write your own. The builder organizes your inputs; it does not assess your expertise or send information to an AI service.
Your reusable draft
This is a structured draft. Check the inputs and the result in the tool you choose.
Keep your next step.
Download your notes, lesson completion and brief. Review the file before sharing it; it can contain what you typed.