Turn a request into conditions you can check
Review specific prompts, targeted revision, deterministic calculations, external instructions and evidence of tool actions.
The five-item count checks stated choices. It is not a live measurement of your prompting speed or workflow productivity.
Name the task, audience, source and format
A request becomes easier to inspect when it specifies the intended output, reader, allowed source and a measurable constraint. For a new fictional exercise, ask for a short welcome note based only on three supplied facts, with missing details flagged. A precise request can guide generation but cannot guarantee compliance. Check the resulting text against each stated condition. Do not convert a request to use only a source into a claim that the tool must therefore have ignored all other influences.
Revise the observed gap rather than the tone alone
If a summary misses an exception that is present in the supplied source, point to that passage and request accurate inclusion. Then inspect the revised version. Asking for a more confident voice does not address the missing condition. In a fresh exercise, mark one omitted date restriction and one omitted eligibility exception, then compare the revision against both. The task remains source-based; inventing a replacement exception from memory would change the evidence rather than repair the summary.
Use a calculation when the result must be exact
For a table containing 56 and 44 only, the exact total is 100. A trusted calculation directly tied to those inputs provides a check of a generated total. A language model can participate in a workflow with calculation tools, but a prose explanation does not prove that a calculation ran. Identify the supplied values, operation and result before checking the claim. This guide illustrates a small invented example and does not claim that language models are incapable of arithmetic.
Keep external content separate from task authority
A page or file can include text that tries to redirect an assistant away from the user’s task. For a benign exercise, place an unrelated slogan instruction inside a fictional document to be summarised. Explain why it is source content rather than permission to change the task. Clear boundaries and technical controls can reduce risk, but a sentence telling a model to ignore attacks is not proof that all prompt injection is prevented. No external instruction is executed by this quiz.
Ask for evidence of an action, not just its narration
An assistant saying it sent a notice does not establish a send operation. The declared capabilities, granted permission and actual record of execution are separate pieces of evidence. If the scenario gives no sending tool, do not invent one to make the statement true. A fresh exercise can compare a draft message, an attempted send and a confirmed delivery record without sending anything. State precisely which stage the supplied evidence supports, including any uncertainty about receipt.
Build a small checkable rehearsal
Choose an invented task with three source facts, one formatting requirement and an explicitly limited tool set. Write the requested output yourself, add one deliberate mistake and make a checklist that would detect it. This rehearses verification without requiring a model, account or private input. If you later try an authorised tool, compare its actual output and execution evidence against the same checklist. Time spent verifying is part of the workflow rather than something a confident answer makes disappear.
Interpret five choices within this practice bank
The section gives a correct count out of five with no productivity multiplier, tool ranking or hiring threshold. Every count includes all worked answers and the other complete guides. A useful next exercise changes the source facts or the tool permission, then asks you to explain what new check is needed. Memorising option positions in this fixed bank is different from demonstrating a new workflow. The website performs no email, publishing or model action as part of the questions.