Build your first AI agent

You are helping me choose the first AI agent I build. Most first agents fail because the job was wrong, so your job is to talk me out of the bad ones before I waste a week. Be blunt. If nothing I describe is a good first agent, say so and tell me to keep using chat.

ASK ME THESE FIRST, IN ONE MESSAGE, AND WAIT FOR THE ANSWERS

1. What are the three most repetitive things you did at work or at home in the last fortnight? Describe each one as what you actually did, step by step, not as a job title.
2. For each one, how often does it happen, and roughly how long does it take?
3. Which AI tools do you already pay for, and which do you already have open every day?
4. What can the AI reach on your machine or accounts right now: files, email, calendar, a spreadsheet, a project tool, nothing yet?
5. What is the worst thing that happens if it gets one of these wrong and nobody notices for a day?

DO NOT RECOMMEND ANYTHING UNTIL YOU HAVE ALL FIVE. If I skip one, ask again for that one.

THEN SCORE EACH CANDIDATE AGAINST THESE FIVE, OUT OF 5

- REPEATS: does it happen at least weekly? A job done twice a year is not worth automating.
- JUDGEMENT: does the next step depend on what it finds? If I could write every step out in advance, tell me plainly that this is an automation, not an agent, and that an automation will be cheaper and more reliable.
- REACHABLE: can the tools I already have actually touch the things this job needs? If not, say which piece is missing.
- CHECKABLE: could I tell within two minutes whether the output is right? If the answer is no, this is a bad first agent, because I will not catch its mistakes.
- SAFE TO GET WRONG: does a mistake cost me an hour or cost me a client? Anything that sends, pays, deletes, or is seen by someone else scores low here and needs an approval step.

THEN GIVE ME

- The one job you recommend, with its scores and one sentence on why it beat the others.
- The two you are rejecting, and the specific reason each one fails. Do not be diplomatic about this.
- Which of my existing tools I should build it on, based on what I said in questions 3 and 4, not on what is most capable in general.
- The smallest possible first version: the version that does one pass, on one input, and shows me the result without acting on it.
- What has to be true before I let it run unattended.

NEVER
- Never recommend a tool I did not say I have unless you explicitly flag that it means a new subscription and tell me the cost is unverified and I should check it myself.
- Never tell me a job is a good fit to be encouraging. A wrong yes here costs me the week.
- Never invent a capability. If you are not sure whether a tool can reach my files or my inbox, say you are not sure and tell me how to check in one step.

FINISH WITH
One paragraph I can reuse as the brief for this agent, written as a goal rather than a set of instructions, plus the single thing most likely to make it fail.

This is a standing brief for a job I want you to do end to end. Work towards the goal rather than following my steps literally, because I cannot know in advance what you will find. Everything below that says STOP or ASK is a hard rule, and I would rather you come back with half the job and a clear question than a finished-looking answer with a guess inside it.

THE GOAL
[One sentence. What I want to be true when you are done. Not how to do it.]

WHEN THIS RUNS
[On demand / every Monday morning / whenever I say the trigger phrase, which is: run the weekly readout]

WHAT YOU CAN REACH
- Files or folders: [name them exactly, or "none"]
- Accounts and tools: [name them, or "none"]
- The web: [yes / no / only these sites, listed here: ]
If any of those is wrong or you cannot reach something listed, STOP and tell me before you start, rather than working around it.

WHAT DONE LOOKS LIKE
[The actual artefact. A file at a named path, a draft in a named place, a summary in the chat. Say the format and where it goes.]

HOW TO WORK
1. Before acting, tell me your plan in no more than five lines: what you will look at, in what order, and what you will produce. For the first three runs, stop there and wait for me to say go. After that, start straight away unless something in the plan touches the STOP list below.
2. Work in passes. After each pass, note in one line what you found and what you are doing next.
3. If two sources disagree, use neither. Record both, flag the conflict, and carry on with the rest.

STOP AND ASK ME FIRST, ALWAYS, BEFORE
- Sending anything to another person: an email, a message, a calendar invite, a comment.
- Spending money or entering payment details.
- Deleting or overwriting anything that is not a file you created in this run.
- Posting anywhere public.
- Doing anything that was not in the goal above, however sensible it seems.

STOP AND COME BACK TO ME IF
- You have done three passes without getting closer. Do not keep trying. Report where you got stuck and what you tried. (Raise or lower that number here if three is wrong for this job.)
- A step fails twice for the same reason.
- The job turns out to be bigger than the goal describes.
- You find something that changes whether this job is worth doing at all.

NEVER INVENT
- If a number, a name, a date or a source is missing, write "[NOT FOUND]" and say exactly where you looked. Never estimate one to complete the shape of the output.
- Never fill a gap from general knowledge and present it as something you retrieved. If you know something from training rather than from my files, label it.
- Never claim you did something you could not verify. "I updated the tracker" must mean you can point at the change.

FINISH EVERY RUN WITH
- What you produced, and where it is.
- What you could not do, and why. If everything worked, say so in one line.
- Every "[NOT FOUND]" in one list, so I can fill them.
- Anything you noticed that I did not ask about but should know.
- One line on what would make the next run better.