# Where are you on your AI journey? An interview guide for AI agents

> Use this page to ask a person fourteen short questions, one at a time, and tell them their level from 1 to 5 and their next three moves. Free, about five minutes, for anyone who uses an AI, whatever the tool.

Updated 2026-10-26. Written from the page code, so it matches what the page does. Page for people: https://both-hands.com/assessment/journey . Index: https://both-hands.com/llms.txt and https://both-hands.com/agent.json . Links use https://both-hands.com/assessment; if you fetched this file from another address, use that one.

## Rules of the interview

1. **One question at a time.** Ask, wait for the answer, then move on. Do not paste the whole list.
2. **Teach in plain words and never quiz.** If a question uses a word they may not know, say what it means in one sentence. Never test them.
3. **Ask only what the person alone knows.** If you are the AI they use, you may offer what you can see of your own setup (for example that you have no saved memory of them) and let them confirm. Otherwise ask. They confirm every answer.
4. **Honest answers.** Tell them to answer "no" as often as it is true. Do not nudge them toward a higher score, and do not answer for them.
5. **Change nothing without a yes.** Do not save, send, connect, install or contact anyone while you do this.
6. **Keep it to this task.** There is no API for you to send answers to. The page has a contact form that sends the person's result to Both Hands, or opens their own email app; it is for the person alone, so never fill it in, press it or write to us for them. Do not share their answers with anyone.
7. **Interview in their language** (English or French are fine). The result wording below is English.
8. **No em dashes or en dashes** in what you write for them.

## The fourteen questions

Reply options are the letters a, b and c in the order shown. Show them the answers in their words, not the points.

### Part 1 of 5: How you use it

What you hand it matters more than how often you use it.

**1.** How often do you use AI for real work?
- a. most days, it's part of how I work
- b. a few times a week
- c. rarely, or not yet

**2.** What do you mostly use it for?
- a. whole pieces of work, from a rough brief to a finished thing
- b. drafting, summarising and questions
- c. quick answers, like a search engine

**3.** How do you ask it for something?
- a. I explain the problem, the limits and what good looks like, then let it suggest
- b. I tell it what I want, then correct it
- c. I type a question and take the first answer

### Part 2 of 5: Whether it knows you

An AI that starts from zero every morning can't get better at working with you.

**4.** Have you written it a standing note about you and how you work (custom instructions, or a file like CLAUDE.md)?
(A note your AI reads every time. ChatGPT calls it custom instructions. The Claude app calls it Instructions for Claude, and Claude Code reads a CLAUDE.md file. Muse keeps what it knows about you in Memory, a dot in its Custom rules, a Grok Bot in its description.)
- a. yes, and I keep it up to date
- b. something short I rarely update
- c. no

**5.** Does it know who you work with, what you're responsible for and how you like things done?
- a. yes, written where it reads it every time
- b. it picks it up in a conversation, then forgets
- c. no

**6.** Does it remember past work, decisions and people from one session to the next?
- a. yes, it has a memory it can search
- b. I paste in notes when I remember
- c. no

### Part 3 of 5: What it can reach

An assistant that can't reach your work can only give advice.

**7.** Is it connected to the tools you work in, like email, calendar, files or chat?
- a. three or more, and I use them every day
- b. one or two
- c. none, I copy and paste

**8.** Does it have saved, reusable ways of doing the jobs you repeat?
(Tools call these skills, custom GPTs, projects, goals or saved prompts.)
- a. several, and I keep improving them
- b. one or two
- c. no, it starts from scratch every time

**9.** If your laptop died today, could you set your assistant up again in an hour?
(Your instructions, saved how-tos and settings, kept somewhere you could restore them from.)
- a. yes
- b. partly
- c. no, it would be gone

### Part 4 of 5: What it does on its own

Going from chat to agent is going from advice to action.

**10.** Can it do things, not just chat: run tasks, work with files, use tools?
(Claude Code, Codex or Cursor work on your files directly. Muse, Dots and Grok Bot act for you in your apps and a browser. A chat window that only answers counts as chat only.)
- a. yes, and I use it that way most days
- b. I've tried an agent that can
- c. no, chat only

**11.** Does it do work while you're away, like scheduled or overnight tasks?
(A summary waiting for you each morning, a weekly check that runs on its own. Some agents call these schedules, routines or goals.)
- a. yes, and I check what it produced
- b. I've tried it once or twice
- c. no

**12.** Have you written down what it must never do on its own, like send, pay, delete or post as you?
- a. yes, written down and enforced
- b. a rule in my head
- c. I hadn't thought about it

### Part 5 of 5: What it gives back

Working more hours because of AI is a warning sign, not a badge.

**13.** When it hands you something, what do you do?
- a. I check it before I act, every time
- b. I usually check
- c. I take it as it is

**14.** Has AI given you time back?
- a. yes, hours a week I can point to
- b. it's changed the work, not the hours
- c. I work more than before

## Score it exactly like the page

Points: answer **a = 2, b = 1, c = 0** for every question. The total is out of 28.

1. **Level by score**:

   - 24 to 28 points: Level 5, Agentic
   - 18 to 23 points: Level 4, Configured
   - 12 to 17 points: Level 3, Agent owner
   - 6 to 11 points: Level 2, Chat user
   - 0 to 5 points: Level 1, Curious

2. **A level never claims what the answers deny.** Check every rule below. If any holds, the level cannot be higher than its cap. With several, use the lowest cap. (Q4 is question 4, and so on; "points" are the a=2, b=1, c=0 values.)

   - If Q10 is exactly 0 points: cap at Level 2, because it can talk but not act yet.
   - If Q10 is exactly 1 point: cap at Level 3, because you have tried an agent but do not use one yet.
   - If Q11 is exactly 0 points: cap at Level 3, because nothing runs while you are away.
   - If Q4 + Q5 + Q6 is 2 points or fewer: cap at Level 3, because it does not know you yet.
   - If Q4 + Q5 + Q6 is exactly 3 points: cap at Level 4, because it only knows a little about you.
   - If Q6 is exactly 0 points: cap at Level 4, because it forgets everything between sessions.
   - If Q11 is exactly 1 point: cap at Level 4, because it has worked on its own only once or twice.
   - If Q12 is exactly 0 points: cap at Level 4, because there are no written rules on what it may never do.
   - If Q12 is exactly 1 point: cap at Level 4, because the rules on what it may never do are only in your head.

   The final level is the lower of the level by score and the lowest cap. If the cap made it lower, say: "Your score alone would put you at Level {by score}. What holds you here: {the reasons of the rules whose cap equals the final level, joined with ", and "}."

3. **The level text**:

   - Level 5, Agentic: Your AI knows how you work, acts on your files and tools, runs jobs while you are away and stays inside rules you wrote down. The next gain is other people: hand your best procedures to a library your colleagues' agents can read.
   - Level 4, Configured: It acts on your work, and you have started teaching it about you. What is missing is usually one of three: a memory that survives the session, a first scheduled job, or a backup you have tested.
   - Level 3, Agent owner: You have an agent that can act, or you have tried one. It does not know you well enough yet to act well. An instructions file and a memory that lasts are the two moves that change the most from here.
     (If Q4 + Q5 + Q6 is 5 or 6 points, use this instead: You have an agent that can act, or you have tried one, and you have taught it how you work. The next gain is letting it do more of the work: your moves below show where.)
   - Level 2, Chat user: You use it, and it helps. The step up is an agent that can act on your files and tools. It goes faster with someone beside you who has done it.
   - Level 1, Curious: Nothing is wrong. Start here: pick one real task this week, hand it the whole problem rather than a question, and see what comes back.

4. **Pillar bars**: for each of the five parts, the points scored and the maximum (How you use it: 6; Whether it knows you: 6; What it can reach: 6; What it does on its own: 6; What it gives back: 4).

5. **Your next three moves**: take every question answered with fewer than 2 points. Sort them by points, lowest first; break ties with this order of question numbers: 10, 4, 6, 7, 8, 11, 12, 9, 5, 3, 2, 1, 13, 14. Take the first three. Each question has one move:

   - Q1: Put it in the first hour of your day for a week. Use is a habit before it is a skill.
   - Q2: Give it one whole piece of work, start to finish, rather than a paragraph of it.
   - Q3: Next time, hand it the problem and the constraints, not the answer you already have in mind. Ask it to propose three ways before it builds one.
   - Q4: Write one page about you: role, who you work with, how you like things done, what it must never do. Put it where your AI reads it first. Our prompt gets your AI to interview you and write it. (Our prompt: https://both-hands.com/assessment/ai)
   - Q5: Add the five people you work with most, and one line each on what they need from you.
   - Q6: Give it a memory that outlives the session. CortexMind is ours; there are others. The test is whether it can find a decision from last month without you pasting it. (CortexMind: https://cortexmind.net)
   - Q7: Connect one tool this week, the one you copy and paste from most. Email or files, usually.
   - Q8: Take the job you did three times this month and write the procedure once, as a skill it can run.
   - Q9: Keep a copy of your instructions, skills and settings somewhere safe: a private git repository if you can, a shared folder if not. Then check you could set it up again from that copy.
   - Q10: Try an agent that can act: Claude Code or Codex on your own files, or a personal agent like Muse, Dots or Grok Bot in your apps. Pick one, spend an afternoon, with someone who has done it beside you if you can.
   - Q11: Give it one job to do while you are away, something with a visible result, and check the result rather than the fact it ran.
   - Q12: Write the no-go list: what it may never send, pay, delete or post in your name without you. Five lines. Then make the rule a mechanism, not a memory: most agents now have approval settings, so set them to ask first.
   - Q13: Before you act on anything it produced, ask it where each claim came from. If it cannot say, it is a guess.
   - Q14: Name the hours you want back, and give it those jobs first. If it is adding work, it is doing the wrong jobs.

   If no question is below 2 points, say: "Nothing on this list needs fixing. Your next gain is other people."

## What to tell them

Give: their level and name ("Level N of 5: Name"), the score out of 28, the level text, the cap line if there is one, the five pillar scores, and the three moves. Keep it short. Then, only if they want it, mention that the prompt at https://both-hands.com/assessment/ai.md interviews them and writes their standing instructions (if you can open that file, you can offer to follow it with them now). Do not contact anyone or send anything for them.

The page's own result also has a "Hand this to your agent" block: a short message, built from the scores only (their level, their strongest and weakest parts and the moves), that the person pastes into their AI so it can pick the next move with them. You do not need to reproduce it, since you are that conversation already. Fit the moves to the person: the page's moves assume nothing about their tool, so if a move does not fit what they use (for example "Try Claude Code or Codex on your own files" for someone who only uses a chat app), say what the move is for and offer the nearest step they can take today. Keep the scores, the level and the cap line exactly as computed.

## Prefill link: open the page with the answers filled in

When you have their answers, you can give them a link that opens the page with those answers already chosen: https://both-hands.com/assessment/journey#fill=PAYLOAD , where PAYLOAD is the base64url of this JSON as UTF-8 bytes, or the JSON itself percent-encoded (use that if you cannot compute base64 reliably; spaces as %20, never +). Put the link on its own line, or in a code block, with nothing after it: a full stop, bracket, asterisk or quote stuck to the end is the usual reason a link fails.

```json
{"v":1,"answers":{"1":"a","2":"b","3":"c","4":"a","5":"b","6":"c","7":"a","8":"b","9":"c","10":"a","11":"b","12":"c","13":"a","14":"b"}}
```

- `v` must be `1`. `answers` is an object keyed by question number ("1" to "14"), each value `a`, `b` or `c` (the letters above; case does not matter). Partial answers are fine: the page opens with those chosen and the rest open.
- Anything else is ignored. Values that are not allowed are dropped and listed to the person on the page. A link longer than 16,000 characters is refused.
- The part of a link after the # sign never reaches a server, so opening it sends nothing anywhere. **Nothing is recorded anywhere**: answers stay in that browser, and the page sends nothing unless the person fills in the contact form at the end and presses its button. Their saved answers on that device are not overwritten.
- Give the link only to the person. It holds their answers.

The link for the sample answers below: https://both-hands.com/assessment/journey#fill=eyJ2IjoxLCJhbnN3ZXJzIjp7IjEiOiJhIiwiMiI6ImIiLCIzIjoiYyIsIjQiOiJhIiwiNSI6ImIiLCI2IjoiYyIsIjciOiJhIiwiOCI6ImIiLCI5IjoiYyIsIjEwIjoiYSIsIjExIjoiYiIsIjEyIjoiYyIsIjEzIjoiYSIsIjE0IjoiYiJ9fQ

## If you drive the page in a browser instead

Prefer the prefill link: it is exact. If you click through the page, each answer is a native radio button inside a fieldset whose legend is the question. The radio for answer a of question 4 is `#q4-a`, and so on (`#q1-a` to `#q14-c`). The result appears below the questions once all fourteen are answered. Never fill in or press the contact form at the end: it sends to Both Hands or opens the person's own email app, and writing to us is their choice.

## Worked examples

Computed from the same data as the page. The tests open the page with these answers and check it agrees.

**Mostly the first answer**: answers aaaaaaaaaaaaaa

- Score 28 of 28. Level by score 5. Final: Level 5 of 5, Agentic.
- Pillars: How you use it 6 of 6; Whether it knows you 6 of 6; What it can reach 6 of 6; What it does on its own 6 of 6; What it gives back 4 of 4.
- Moves: none

**A mixed set**: answers abcabcabcabcab

- Score 15 of 28. Level by score 3. Final: Level 3 of 5, Agent owner.
- Pillars: How you use it 3 of 6; Whether it knows you 3 of 6; What it can reach 3 of 6; What it does on its own 3 of 6; What it gives back 3 of 4.
- Moves: Q6: Give it a memory that outlives the session. CortexMind is ours; there are others. The test is whether it can find a decision from last month without you pasting it. | Q12: Write the no-go list: what it may never send, pay, delete or post in your name without you. Five lines. Then make the rule a mechanism, not a memory: most agents now have approval settings, so set them to ask first. | Q9: Keep a copy of your instructions, skills and settings somewhere safe: a private git repository if you can, a shared folder if not. Then check you could set it up again from that copy.

**Mostly the last answer**: answers cccccccccccccc

- Score 0 of 28. Level by score 1. Final: Level 1 of 5, Curious.
- Pillars: How you use it 0 of 6; Whether it knows you 0 of 6; What it can reach 0 of 6; What it does on its own 0 of 6; What it gives back 0 of 4.
- Moves: Q10: Try an agent that can act: Claude Code or Codex on your own files, or a personal agent like Muse, Dots or Grok Bot in your apps. Pick one, spend an afternoon, with someone who has done it beside you if you can. | Q4: Write one page about you: role, who you work with, how you like things done, what it must never do. Put it where your AI reads it first. Our prompt gets your AI to interview you and write it. | Q6: Give it a memory that outlives the session. CortexMind is ours; there are others. The test is whether it can find a decision from last month without you pasting it.