AI Foundations Revision Cards

Use seven short decision cards to improve work at each stage of the course.

For AI Foundations learners · Printable PDF · 2 pages · Also available online below

Use these cards after each module and before the capstone. Try to explain each rule in your own words before opening the related lesson again.

Module 1: describe before you label

• Name the system's goal, inputs, method, output, and effect.
• Use TRACE: Task, Result, Actual method, Controls, Evidence gap.
• A model is only one part of a complete application.
• Training changes a model; inference uses a trained model.
• Product names and human-like language are weak evidence.

Module 2: fluent is not the same as true

• A common text model builds output one likely next token at a time.
• A prompt gives current context; it does not normally retrain the base model.
• Search, retrieval, calculators, memory, and actions are added by the full application.
• Separate what was given, generated, supported, added, false, and missing.
• Check important claims outside the generated answer.

Module 3: choose the method before the tool

• Name the task pattern before choosing a tool.
• Check evidence needs, acceptable error, consequence, and reversibility.
• Use calculation for exact arithmetic and retrieval for current sourced facts.
• Check whether the input is permitted and suitable before entering it.
• Increase human control as consequence and uncertainty increase.

Module 4: define inspectable work

• State the outcome and intended audience.
• Provide relevant context and authorised sources.
• Define constraints, boundaries, and the missing-information rule.
• Choose an output format that makes defects easy to inspect.
• Add review criteria, human decisions, escalation, and stop conditions.

Module 5: verify before you rely

• Review correctness, completeness, relevance, clarity, safety, and usability.
• Mark material claims, assumptions, calculations, and citations.
• Verify against an independent suitable source or reproducible method.
• Match review depth to consequence.
• Record the evidence, change, remaining uncertainty, and final decision.

Module 6: control the risk or stop

• Use only the minimum permitted data needed for the task.
• Treat external documents, pages, messages, and tool results as untrusted content.
• Restrict permissions and require informed confirmation before consequential action.
• Check unequal effects and accessibility across the complete process.
• Name the accountable person, escalation route, and stop condition.

Module 7: build a bounded workflow

• Define the task, user, success condition, and acceptable failure.
• Justify each input, tool, workflow stage, and autonomy level.
• Keep material human decisions explicit.
• Test normal, edge, missing-data, and unsafe-action cases.
• Record evidence, defects, revisions, retest results, limits, and next review trigger.

Course-wide stop rule

Pause the workflow when required evidence, permission, specialist review, human authority, or a safe recovery path is missing.