A practical guide to learning with AI, from your first attempt to independent application.
Follow twelve lessons across four stages: Frame, Practise, Verify and Apply. Each lesson includes a worked example, a decision and two practice tasks. Choose a project and use its workbook checklist to collect your evidence.
Use your browser’s Print command to save a PDF or print this guide. Expand answer sections before printing if you want the reasoning included. All supplied project situations and data are fictional. Progress is self-assessed; this guide does not provide automatic grading or connect to an AI service.
Frame · 7 minutes
Start with something you want to do, not a topic you want to consume. A useful learning outcome names an action, a context and a standard. This gives your AI assistant a coaching target and gives you a way to recognise progress.
Suppose you want to “learn spreadsheets.” That could mean hundreds of unrelated skills. “Summarise one week of orders and explain which product sold most” is small enough to attempt today. Choose an outcome slightly beyond your current ability, with a result you can inspect. You can expand the scope after the first successful attempt. In this guide, vibe learning means a practical loop: set a direction, attempt, use AI for targeted support, check the result and apply the idea independently.
A community volunteer needs to plan a workshop. Instead of “learn event planning,” she chooses: “Create a 45-minute agenda for 12 beginners, with clear instructions and a five-minute buffer.” The agenda is an artifact; the timing and instructions are quality criteria.
Write an action, a situation and two quality criteria for something you want to learn this week.
Which goal is easiest to assess?
Write a 120-word update with a clear decision, reason and next step.
A concrete output and criteria let you inspect your work.
Adapt your goal for a person who has half as much time. What would you reduce without losing its purpose?
Frame · 8 minutes
Make a short unaided attempt before asking for an explanation. A baseline reveals specific gaps and gives you something to compare with later. It is a diagnostic, not a judgment about your ability.
Set a small time limit and record what you can do without notes or AI. Separate “I cannot recall the steps” from “I can do the steps but cannot explain why.” Those gaps need different practice. Do not compare an easy baseline with a much easier final task and call the difference mastery. Keep the later check similar in difficulty but change the details so you cannot simply repeat the answer.
Before learning chart interpretation, a learner writes three observations about a bar chart. Two are accurate; the third mistakes totals for percentages. The next session focuses on labels and denominators rather than reviewing every chart type.
Spend three minutes attempting your outcome. Record one thing you can already do, one uncertainty and one missing skill.
Your first attempt is mostly wrong. What is the most useful response?
Save it, identify one gap and choose a targeted practice task.
The original attempt helps you see what changed and why.
Design a second baseline question that tests the same skill in a different situation.
Frame · 7 minutes
Break the outcome into a few abilities you can practise. Put prerequisites before tasks that depend on them, and stop the plan from growing into a course you never start.
Ask what a successful person must notice, decide and produce. Then identify the first bottleneck. For a short report, the route might be selecting a claim, choosing evidence and writing a recommendation. You do not need to study every writing technique before producing a first report. Use AI to suggest a route, but trim steps that do not serve the chosen outcome. After each attempt, revise the route based on evidence rather than loyalty to the original plan.
To analyse workshop feedback, a learner first distinguishes a count from a percentage, then calculates both, then writes a cautious summary. Formatting the final slide comes afterward because a beautiful slide cannot repair a misleading calculation.
List three abilities your outcome requires. Mark the first bottleneck and a ten-minute practice task for it.
Your AI-generated plan has 40 topics. What should you do first?
Keep the few prerequisites needed for your first useful output.
A short route makes the next attempt manageable.
Which part of your route could you remove and still produce a useful first version?
Practise · 9 minutes
Give AI your goal, your attempt and a specific request for support. Ask it to wait for you to think. The key is not a magical prompt; it is a turn-taking pattern that preserves your decisions.
A coaching request can say: “Here is my attempt. Ask one diagnostic question, then wait. If I am stuck, give the smallest useful hint.” This reduces the temptation to accept a complete answer before trying. AI can still be wrong, overhelpful or unclear. If it gives away the solution, close it and attempt a fresh example. Do not share private work details when a fictional or anonymised example would teach the same skill.
Weak request: “Write the client email for me.” Learning request: “I drafted this fictional update. Does the requested decision appear clearly? Ask me one question before suggesting a rewrite.” The learner then revises the email and explains the change.
Write a coaching prompt containing your outcome, your current attempt, one difficulty and a request to wait for your response.
The assistant immediately supplies the full answer. What keeps the session focused on learning?
Move to a fresh example, try first and request only a hint.
A fresh attempt restores the opportunity to make your own decisions.
Rewrite your coaching prompt for a task where you need help spotting an error rather than learning a new concept.
Practise · 8 minutes
Start with the smallest support that helps you move. A useful ladder is: clarify the goal, ask a guiding question, reveal one step, show an analogous example, and only then inspect a full worked solution.
Different problems need different amounts of support. Remaining stuck indefinitely is not the goal. If a small hint does not unlock an attempt, move up a level and record the support you used. After a worked solution, explain the reasoning and solve another problem with less help. Counting a heavily supported answer as independent success makes it harder to decide what to practise next.
A learner cannot calculate a percentage. First hint: “What is the whole?” Second hint: “Divide the part by the whole.” Analogous example: “Three out of twelve is 3 ÷ 12 × 100.” The learner then calculates seven out of twenty without copying the example.
Write three hints for your current task: a question, one partial step and an analogous example. Keep the final answer separate.
You needed a full worked solution. What should happen next?
Try a new example with reduced support and record what you can do.
Fading support makes the next attempt informative.
Choose a hint you could remove from your next attempt. What evidence would show that you no longer need it?
Practise · 9 minutes
Turn a mistake into a specific practice target. Identify the first point where your reasoning diverged, explain the cause and test a correction on a new example.
“I am bad at this” is not a useful diagnosis. “I divided by respondents instead of all invitees” tells you what to inspect. Classify the difficulty: missing knowledge, mistaken assumption, execution slip or misunderstood instructions. A slip may need a checklist; a conceptual misunderstanding needs examples and explanations. Ask AI for possible causes, then use your own work to decide which cause fits.
A learner reports 30% growth when a count rises from 20 to 30. The increase is 10, so growth is 10 ÷ 20 = 50%. The learner confused the new total with the percentage increase. Practice should distinguish starting value, change and final value.
Record one mistake, its likely cause, a corrected explanation and a new example that tests the correction.
Which reflection is most useful?
I used the final value as the denominator; next time I will label the starting value before calculating.
This identifies a cause and a concrete corrective action.
How would you tell whether the same mistake came from a conceptual gap or a rushed execution slip?
Verify · 10 minutes
Treat an AI answer as something to investigate. Pick claims that matter and check them with an original source, a reproducible calculation or an independent test.
A citation is not enough: open the source and see whether it supports the exact claim under the stated conditions. A calculation should be reproducible from the inputs. A software suggestion should be tested on a small, non-sensitive example. Asking the same assistant “Are you sure?” may reveal a problem but is not independent verification. Keep a note of what you checked, what it showed and what is still uncertain.
An AI summary says 80% of invited residents supported a proposal. In the fictional data, eight of ten respondents supported it, but fifty people were invited. Support among respondents is 80%; confirmed supporters among invitees are 16%. Neither establishes what all residents believe.
Choose one important claim from your project. Record a check, the result and a limitation.
A source is real but discusses a different population. What should you do?
Narrow the claim or seek evidence matching your situation.
Matching the claim to the evidence avoids overstating the result.
Find a statement in your own project that sounds more certain than your evidence permits. Rewrite it.
Verify · 8 minutes
When there is no single right answer, define what matters before choosing. Use explicit criteria, constraints and tradeoffs rather than accepting the most confidently presented option.
A recommendation depends on the situation. A workshop activity that works for six experienced adults may fail for thirty beginners. Name non-negotiable constraints first, then compare plausible alternatives. A score table can organise judgment, but weights are choices, not objective truth. Ask what change in the conditions would reverse your recommendation.
Two fictional workshop plans both fit a 45-minute slot. One offers more content but no practice; the other covers fewer ideas with a short independent task. If the outcome is demonstrated skill, the second plan may fit better. If the purpose changes to a broad introduction, the comparison changes.
Compare two possible project approaches using three criteria. State one tradeoff and one condition that would change your choice.
What makes a recommendation easier to review?
Explicit criteria, relevant evidence and a condition that could change the decision.
Readers can inspect the reasoning and challenge assumptions.
Change one constraint in your project. Does your preferred option still fit?
Verify · 9 minutes
Transfer means applying an idea when surface details change. Try a new context, a new constraint or a different representation, and explain why the same principle still applies.
A near-transfer task changes small details; a farther-transfer task changes the setting. Begin with a manageable change, then increase the distance. If performance drops, locate which part of the idea did not carry over. Do not assume success on one example means the skill will transfer everywhere. Use the task requirements to decide which variations matter.
After calculating a survey response rate, calculate an event attendance rate using registrations as the denominator. Then ask whether attendance rate and satisfaction rate answer the same question. The arithmetic transfers, while the meaning of the denominator requires renewed attention.
Create one near-transfer and one farther-transfer task for your project skill. Attempt the near-transfer task without AI.
Which is a better transfer check after learning to summarise feedback?
Summarise a new feedback set and explain what you excluded and why.
New evidence and an explained selection reveal more about the underlying skill.
What changed between your original task and the new one? What stayed the same?
Apply · 10 minutes
Make a result another person can inspect: an email, analysis, plan or small working design. Include the reasoning needed to understand important choices, not every step of your learning history.
An artifact connects learning to use. Before polishing it, test whether it meets the original outcome. For a data summary, check calculations and limitations; for an email, check the reader’s requested action; for a workshop, check timing and participation. Use AI as a reviewer against your criteria, then decide which suggestions to adopt. Keep a short change log explaining at least one revision.
A volunteer creates a one-page event brief containing an agenda, materials list and contingency. A reviewer notices that setup time was omitted. The volunteer revises the schedule and explains which activity was shortened and why.
Create a first version of your project artifact. Review it against two criteria and describe one evidence-based revision.
What should you check before improving visual polish?
Whether the artifact fulfils its purpose and its important claims are correct.
Purpose and accuracy come before presentation refinements.
Imagine someone using your artifact without you present. What instruction or assumption needs to be clearer?
Apply · 9 minutes
A useful independence check is a short explanation without your notes. State the purpose, a key decision, the evidence for it and a limitation. Then respond to a changed condition.
This is not about delivering a perfect speech. It is a way to find parts of the project you can operate but do not yet understand. If you cannot explain a calculation, rebuild it from the inputs. If you cannot justify a recommendation, revisit the criteria. Ask a partner or your AI assistant to challenge one assumption, but answer before requesting model feedback.
A learner explains why a survey summary says “among respondents.” When asked whether the result represents everyone invited, she explains nonresponse and declines to make an unsupported claim. This demonstrates judgment as well as arithmetic.
Write a 90-second explanation of your artifact without looking at it. Include one choice, one check and one limitation.
You cannot explain a step you copied from AI. What is the next useful action?
Reconstruct the step on a small example, then explain it again unaided.
A small example helps connect the procedure to its purpose.
Invent a skeptical but fair question about your artifact. Answer it using evidence rather than reassurance.
Apply · 8 minutes
Finish a project with a review plan rather than a claim of permanent mastery. Keep evidence of what you can do, choose a later unaided check and decide which support to remove next.
Completion records participation. Capability depends on performance on relevant tasks over time. Compare your baseline with your final attempt, noting differences in task difficulty and help received. Choose a short review task that checks the central skill. If you struggle later, use the smallest support that restores progress and repeat a fresh example. A useful portfolio shows attempts, corrections and independent work, not only polished outputs.
After producing a clear update email, a learner schedules a later task: write a different update in ten minutes without AI. She will check whether the decision, reason and next action are explicit. If the structure holds but the wording is weak, her next practice targets concision.
Name your next unaided task, a review date, success criteria and the support you will avoid unless you get stuck.
Which statement is best supported by finishing this guide?
I have completed a learning process and can now test which skills transfer independently.
A clear next check keeps confidence connected to evidence.
Write a new project brief that uses one skill from this quest in a different context.
A neighbourhood workshop invited 50 people. Twenty attended. Ten attendees completed a feedback form: eight rated it useful, two did not. Six respondents asked for a beginner follow-up. Prepare a short recommendation for the next event.
Invited: 50. Attended: 20. Respondents: 10. Useful: 8. Not useful: 2. Requested beginner follow-up: 6. These are fictional practice data.
A 150-word summary with three labelled percentages, a recommendation and two limitations.
Correct denominators; separate respondents from attendees and invitees; avoid claiming that nonrespondents share the same views; make a proportionate recommendation.
Attendance was 20/50 = 40%. Response among attendees was 10/20 = 50%. Eight of ten respondents rated the event useful: 80% of respondents, not 80% of everyone invited. A small beginner pilot is plausible because six respondents requested it, but demand among everyone else is unknown.
Next time, 40 people attend and 12 respond. Nine respondents find it useful. Compare response rates and explain why “satisfaction fell” needs qualification.
Your team’s draft is ready, but review is delayed. The fictional launch is next Monday. You need a colleague to choose by Thursday at noon: reduce scope and keep the date, or keep full scope and move launch to Wednesday.
The draft is ready. Review needs two working days. Reduced scope removes the optional FAQ. Full scope includes it. No decision has been made. All names and dates may be fictionalised.
An email under 140 words with a subject, two options, your recommendation and a clear decision deadline.
Do not invent approval or progress; distinguish facts from recommendations; make the requested decision explicit; use a respectful tone.
Subject: Decision needed Thursday: launch scope. The draft is ready, and review requires two working days. We can keep Monday’s launch by removing the optional FAQ, or include the FAQ and move launch to Wednesday. I recommend the smaller Monday release if the FAQ is not essential. Please confirm your preferred option by Thursday at noon so we can schedule review.
The colleague replies that the FAQ is essential for first-time users. Revise your recommendation and explain which criterion changed.
Plan a 45-minute session for 12 beginners learning to write a useful AI coaching prompt. You have one shared screen, paper and no guarantee that everyone has an AI account.
45 minutes total. 12 beginners. One shared screen. Paper available. Individual AI accounts are optional. Outcome: each learner writes and revises one coaching prompt.
A timed agenda, one fictional practice task, instructions for paired feedback and a no-account fallback.
Agenda totals no more than 45 minutes; every learner makes an attempt; feedback uses explicit criteria; the plan works without individual accounts.
5 min: introduce the outcome. 7 min: compare two sample prompts. 10 min: write a first attempt on paper. 10 min: paired feedback on goal, attempt and request for a hint. 8 min: revise and explain one change. 5 min: independent exit task. Total: 45 minutes. The facilitator can demonstrate one example on the shared screen.
The session starts eight minutes late. Keep the independent exit task and learner practice. Decide what to shorten and justify the tradeoff.
Your club receives event sign-ups in messages, copies names into a sheet and sends reminders manually. Occasionally names are duplicated or a reminder is missed. Design a clearer process before asking AI for automation ideas.
One organiser handles sign-ups. A participant may send more than one message. Required information: name and chosen session. Only the organiser needs the contact details. No real participant data is needed for this exercise.
A five-step workflow, three fictional test cases and a checklist for duplicates, missing information and reminders.
Identify a single source of truth; show how duplicate and incomplete entries are handled; avoid exposing participant details unnecessarily; keep a manual fallback.
1. Receive a sign-up. 2. Check the existing list for that person and session. 3. Request missing details before marking confirmed. 4. Record confirmation and reminder status in the same list. 5. Send reminders only to confirmed participants and record completion. Test a new entry, a duplicate and an entry with no session.
A participant changes sessions after the reminder is sent. Update the workflow so old and new details cannot silently disagree.