AI tools are leverage, and leverage works in both directions. Used well, they speed up drafting, learning, and planning. Used carelessly, they make confident mistakes faster. This area is about the workflows that keep the upside and limit the downside.
Anchor method
Agentic workflows and the extended mind
Treat AI and agents as an extension of how you think: you set the goal and the judgment, the agent carries the steps, and you check the work.
In "The Extended Mind" (1998), philosophers Andy Clark and David Chalmers argued that thinking doesn't stop at the skull: a tool you rely on consistently, like a well-kept notebook, can become part of how you think. AI agents push this further, because they can carry out multi-step work for you. The idea only pays off if you stay the one who sets the goals and judges the results.
Method
Write the goal in one sentence, plus what "done" looks like.
Break it into steps. Decide which steps the agent does and which ones you keep.
Give the agent the context and tools it needs, and nothing private that it doesn't need.
Add checkpoints: the agent stops and shows you its work before anything is sent, bought, or published.
Keep a running notes file that the agent reads and updates, so your extended mind has a memory you can inspect.
Try it
Goal: build our family's first-week school checklist.
Steps: 1) list what the school's website asks for, with links; 2) draft the checklist; 3) stop and show me before contacting anyone.
Rules: cite the page for every item, mark anything you're unsure about, and don't use our names or ID numbers.
Published prompting guides from AI developers give the same core advice: be specific, give the model the context it can't guess, and say what format you want. R-C-O is a 3-block way to remember that.
Method
Role: who should the AI act as? "Act as a careful editor."
Context: who is it for, what is the goal, what are the constraints?
Output: format, length, tone, and what to avoid.
Then verify every fact yourself before you use it.
Try it
Role: Act as a careful editor.
Context: This is a 200-word note to my child's teacher. English is my second language. I want it to be polite and clear.
Output: Return the edited note, then a bullet list of what you changed and why. Do not add new facts.
Generative AI can produce false content that sounds confident. The US National Institute of Standards and Technology lists this (it calls it "confabulation") among the key risks of generative AI. A field experiment with management consultants found that on a task just outside what the AI did well, people using it were less likely to reach the correct answer than people working without it.
Method
Treat every number, name, date, and citation from an AI as a claim to check.
Ask the tool for its sources, then open them. If the source doesn't exist or doesn't say that, drop the claim.
For anything about tax, health, or law, go to the official source directly.
A large field experiment in high school math (Bastani et al., PNAS) found that students given unrestricted access to an AI chatbot did better on practice problems but worse on a later exam taken without it. A version designed as a tutor, with guardrails that gave hints instead of answers, largely avoided that harm.
Method
Tell the AI not to give the final answer.
Ask it to give one question or hint at a time.
Explain your reasoning back to it. Explaining is where the learning sticks.
Try it
Don't give me the answer. Ask me one question at a time until I solve it. If I'm stuck twice, give me a small hint, not the solution.
Whatever you paste into an online AI tool leaves your device and is handled under that provider's policies. NIST's generative AI profile lists data privacy, including leakage of personal information, as a core risk. The safest data is data you never share.
Method
Never paste ID numbers (SIN, SSN), passwords, bank or card numbers, or full medical records.
Replace names and details with placeholders: "[Child], age 9".
Check the tool's data and privacy settings before using it for work.