A path, a lesson a day
Five-minute lessons that build on each other. Spot the invented figure, fix the line that caused it, match the failure to its fix. Mistakes come back until you get them right — and again right before you'd forget.
Free while in early access
Get good at AI. Five minutes a day.
Short, playable lessons that make you measurably better at directing AI at work — then put it to work on the jobs you repeat every week.
- 48bite-sized lessons
- 16real work cases
- 10–160adaptive test score
Try one · Lesson 1
This prompt has a role, a task and a format. What's missing?
"You are a B2B marketing analyst. Summarize the attached Q3 churn report in five bullets."
How it works
Practice daily. Prove it. Get the time back.
A score that proves it
An adaptive 20-minute test, in the style of the Duolingo English Test: one score from 10 to 160 and four subscores. Retake it every two weeks and watch the number move.
Hours back on real work
Every section ends on your own work: the status update, research synthesis or report you do every week, turned into a procedure that runs — timed, checked, and counted in hours saved.
Streaks that count practice, not visits
XP only for right answers and real work
Weekly leagues, and leagues for your team
Review aimed at your weakest skills
What you'll learn
Four sections. 16 units. One skill: directing AI well.
Section 1
Say what you mean
- Anatomy of an instruction
- Context that matters
- Format, length and audience
- Delegating, not abdicating
Section 2
Ground it
- Sources only
- Numbers you can defend
- Synthesis from messy notes
- When the input talks back
Section 3
Check it
- Spot the invention
- Checks that catch things
- How sure is it, really?
- Reviewing like an editor
Section 4
Make it a system
- Specs that run
- Breaking work into steps
- What never goes in a prompt
- From one-off to procedure
If you use AI at work
Sound familiar?
You retype the same instructions into ChatGPT every Monday, slightly differently each time.
One written instruction you reuse, with the sources and the quality bar attached to it.
You skim the output, it reads well, and you send it. You are not really sure what it got wrong.
A check you wrote runs against every draft, and flags the claims nothing in your material supports.
Somebody asks how you made it and the honest answer is: I fiddled with it until it looked right.
A procedure on a page. You can hand it over, improve it, or run it in ten minutes.
Your first real job, about half an hour
You walk away with three things
The finished document
A real deliverable — the status update, the synthesis, the report — with every claim marked against the material it came from, so you know what is safe to send before you send it.
The procedure that produced it
Five lines you keep: what to attach, what to ask, what done means, what to check, and what never goes in. Set a cadence and it is waiting for you next week, on next week’s material.
Proof it’s working
Your score before and after the fix, then every live run timed against how long the job takes you by hand — with a running total of time saved and problems caught before they went out.
Start with whichever costs you the most time
- A status update
- A research synthesis
- A report for executives
- A spec or a brief
- A digest someone else reads
- An analysis of numbers
Real work, step by step
You’ll do it badly first. That’s the method.
On real work there’s nothing to study first. You attempt the job, and whatever goes wrong picks the lesson you get.
- 01
Cold run
about 5 minutesDo the job with no help at all.
It finds your real edges in five minutes. Guessing at them takes weeks — and the gap it finds is what picks the one lesson you get next.
- 02
Repair
Your own sentences, marked where nothing supports them.
A number tells you how you did. A mark under your own words tells you what to change — and each one links to the single lesson that closes it.
- 03
Warm run
The same job again, with hints, a coach and a worked answer.
This is where the document actually gets made, and where you find out whether the fix took. Revise and resubmit as often as you like.
- 04
Keep it
10 secondsSave the procedure and say when it comes round.
This is the part you still have next Monday. Everything above it is proof the procedure works; this is what makes it come back.
What “see where it broke” means
Marked, not scored
A number tells you how you did. A rule under your own sentence tells you what to change — and which claim you should not have sent.
The team resolved most blockers this week1 and the migration is nearly done2.
Checkout latency is down to 310ms at p95, from 890ms3.
- 1. Nothing in your material says how many. Two of seven is “most” to nobody.
- 2. No date. “Nearly done” is what it said last week too.
- 3. Both figures trace to the latency dashboard. Send this one.
What a procedure is made of
Five parts. Only one of them is a prompt.
Prompt craft is the part that depreciates — every model generation absorbs more of it. The other four don’t get easier, and they’re where the work actually goes wrong.
- Sources
What gets attached every time — and what deliberately doesn't?
Without this you paste whatever is nearest and the model answers from nothing.
- Instruction
The prompt you send, every time.
A vague ask gets a fluent answer in the wrong shape, and you redo it by hand.
- Done
What must be true of the output before you'd accept it?
Written after you look, this is just a description of what you got.
- Check
What do you verify before any of this leaves the building?
Skip it and the invented number goes out with your name on it.
- Owner
Who is accountable, and what never goes in?
Nobody owns it, so nobody catches it — and confidential text leaks by habit.
01
02
03
04
05
Beyond the daily path
7 deeper paths and 36 longer lessons
When part of a real run goes wrong, the longer lesson that fixes that part is handed to you — about eight minutes, then straight back to the work.
- AI Foundations for Professionals
- Prompt Craft: From Vague to Precise
- Verify & Trust AI Output
- Automate Your Week
- Directing AI Work
- AI-Powered Product Thinking
- AI for Marketers
Plus guided walkthroughs, on real material
8 screens · 14 min
Raw notes to product bets
Three contradictory user interviews and forty minutes before a roadmap review. Produce a memo where every claim traces back to a quote.
6 screens · 12 min
Audit a shaky report
An AI-drafted investor update with invented numbers in it. Find them, prove they are wrong, and rewrite it before it reaches the board.
5 screens · 10 min
Weekly status in ten minutes
Scattered team updates into a status email your VP will actually read, ending with the prompt you reuse every Monday after.
Looking at this for a team? Team groups are free: a shared league, and a dashboard of who practised and what their real work saved. See how teams work.
FAQ
Questions, answered
Your first lesson is five minutes away
Start the path, find your level with a five-minute placement, and see your score move.
Start learning