How fluent are you with AI? Sixty seconds will tell.
Eight situations from real work with ChatGPT, Copilot, Claude or Gemini. Pick the best move in each. At the end: a profile on the four skills of AI fluency, the level where a programme would start, and a debrief written for you by Claude.
Eight situations. Pick the best move in each.
The clock starts on the first question and pauses while you read each explanation. When it runs out, nothing happens: the test carries on.
Answers stay in this tab. The debrief request carries the answer pattern and the role, nothing else.
A colleague types "Write a summary of our quarterly operations" and gets something generic and unusable. What is the first thing to fix?
Best move: B
The model did not get smarter between a weak prompt and a strong one. The prompt did. Audience, source and format are the parts most often left out, and they are free to add.
A two-hour chat was going well. Now the assistant ignores the format agreed at the start and only answers the last message. Best move?
Best move: C
Context is a budget. A long chat gets summarised behind the scenes and early instructions lose force. Repeating them only extends a degraded session. Summarise, restart, and save what should persist.
A draft for a client reads "adoption rose 63%, up from 41%". No source is given. What does that precision tell you?
Best move: B
Precision is the tell. Invented statistics, dates and quotes read exactly like verified ones, and the tool will confirm them with the same confidence. Trace the claim or drop it.
The assistant produced a forty-line cost table for a board paper, with a total at the bottom. Safest route?
Best move: B
Numbers that matter are computed, not written. A language model writes plausible arithmetic. A spreadsheet or code gives arithmetic that can be checked and re-run.
An HR assistant wants to paste a file of employee names and salaries into a chat tool to sort it. Incognito mode is on. Fine?
Best move: C
Incognito keeps a chat out of history and memory. It does not change where the data goes or who can access it. Classify data before uploading: names plus salaries stay out until an approved path exists.
A colleague shares an add-on found on a public forum that formats reports perfectly, and vouches for it. Enable it?
Best move: C
An add-on is software. It inherits whatever the session can reach: files, connectors, mail. Source and reach decide, not a colleague's word, and a blanket ban just pushes people to personal accounts.
The legal team wants AI in contract review. Which split is right?
Best move: B
Drafting quality is not a licence to delegate the decision. A team that let AI approve low-risk clauses missed an obligation and pulled the tool. A team that kept a lawyer on every change halved its review time.
AI drafted the client newsletter well three months in a row. The team now wants it to send the newsletter to the client list on its own. Agree?
Best move: C
Handing a step to AI because the previous one went well is a classic error. Judge each step on reversibility, stakes and accountability. Sending to clients fails all three, so a named person owns that step.
What the test measures.
Four skills, two situations each. They carry across ChatGPT, Microsoft 365 Copilot, Claude and Gemini.
Asking
Giving the tool what it needs: context, task, constraints, format. And knowing when a chat has run out of room.
Checking
Treating output as a draft. Tracing claims to sources, computing numbers instead of reading them.
Protecting
Classifying data before it goes in, and judging add-ons by who published them and what they can reach.
Delegating
Deciding step by step what AI does, what a person does, and who signs.
Behind the test.
The situations follow the AI fluency framework behind Anthropic's Claude Certified Associate curriculum, the certification Astra Via's founder holds. Its four competencies, description, discernment, diligence and delegation, appear here as asking, checking, protecting and delegating.
The traps are the ones that catch experienced users as often as beginners: the better model that never fixes a thin prompt, the long chat that quietly forgets, the precise figure nobody sourced, the step delegated because the previous one went well.
Where a programme starts.
Starting out
The tool is in use, the method is not yet. AI Essentials builds the habits in one day, on the team's own work.
See AI EssentialsRegular user
Good instincts, with a few traps still open. AI Practitioner turns habits into a team workspace, saved procedures and scheduled work.
See AI PractitionerOperator
The judgement is there. The next step is building: AI Builder, or helping the rest of the team get here through AI Practitioner.
See AI BuilderQuestions.
Is anything stored?
No. Answers stay in the browser. The request for the debrief carries the answer pattern and the optional role, nothing else, and Astra Via keeps no copy of it.
Which tools does the test apply to?
All of them. The situations are about habits, not buttons: ChatGPT and ChatGPT Enterprise, Microsoft 365 Copilot, Claude and Gemini all reward the same moves and punish the same traps.
Why sixty seconds?
Each situation takes a few seconds to answer when the habit is there. The timer is a prompt, not a penalty: when it runs out, the test carries on and the result still counts.
Can a whole team take it?
Yes, the link is open. A team's results are a good starting point for the call that precedes a programme: which level fits most people, and which traps the team rules should close first.
Where does the debrief come from?
From Claude, Anthropic's model, given the answer pattern and the role. It is generated for each result and follows the same house rules as the rest of this site: no prices, no invented figures, one concrete habit to change.