Academy track
Getting started with AI
One job named, one agent built to do it, one schedule it runs on, and a written rule for when to stop trusting it.
- 6 steps
- 4 hours 45 minutes in total
- A check you run at the end of every step
Step one is open to everyone, in full. No account, no card.
The other 5 steps are included in Membership, £79 a month.See Membership
By the end of this track you will have
- Step 1One job written on one line: the input, the output, the owner, and the minutes it costs every week.
- Step 2One agent that reads something, checks it against a rule, and hands you a list. Run by hand, three times, with the results written down.
- Step 3A context file your agent reads before every run, and one rule in it that stops the agent inventing an answer.
- Step 4An agent that lists its sources before its answer, and a check you can run in under a minute to catch it citing nothing.
- Step 5The job running on a schedule, writing to a run log, and an off switch you have tested by switching it off.
- Step 6A written trust ledger for your agent: what it is allowed to do, how you would find out it was wrong, and the date you next check.
The running example
Every worked example follows one invented case from the first step to the last: a four-person lettings agency.
Who it is for
You have tried AI, it impressed you for an afternoon, and a month later nothing of yours runs on it.
Who it is not for
Not for you if an agent of yours already runs on a schedule and keeps a log. Start at Give your AI the keys.
The steps
Each one ends with a check you run yourself. The box is yours to tick when it passes.
0 of 6 steps ticked, kept in this browser
- Step 1.Pick the job before you pick the tool45 minutesOpen to everyone
Done means
You can open the source named in your job line and then open the destination named in it, in under a minute each, without asking anyone where they are. The line also names one owner, one interval and one minutes figure. If any of those five is missing or woolly, the line is not finished.
- Step 2.Your first agent should do something boring60 minutes
Done means
Three logged runs. Two on real input, one on input you broke on purpose. The broken run flags the break you planted, and the clean run returns exactly NO MISMATCHES FOUND with no extra words.
- Step 3.Give it the context it cannot guess50 minutes
Done means
You ask for that missing fact and the agent returns MISSING and names what it needed. If it produces a number, a date, or a confident sentence instead, your file is not finished. Add the never-assume line the failure exposed, and ask again.
- Step 4.Make it show its working40 minutes
Done means
On the trap run the agent outputs MISSING and names what it needed. On the stale-source run it states the source name and its date, and the date is old enough that you can see the problem without asking. If either run produces a confident answer with no source line, the working block is not being obeyed. Move it to the very top of the prompt and run again.
- Step 5.Put it somewhere it runs without you60 minutes
Done means
You check
runs.logafter the next scheduled run time and there is no new line. If a new line appeared, you did not find the switch, and everything you believe about this task being under control is wrong. Go and find the real flag before you go any further. - Step 6.Know when to stop trusting it30 minutes
Done means
trust.mdexists with all six lines filled in and one dated spot check underneath, recording agreement or a defect. A ledger with no check in it is a plan, not a control.
I keep this tick in this browser and send it nowhere. It goes when you close the browser.
Included in Membership
The rest of this track is included in Membership.
Step one is open to everyone, in full. Every step after it opens on a member account.
Included in Membership — £79/month.
What comes next
Your agent reads things now. Give your AI the keys is how it touches a live system without wrecking it.
See all the tracks