“Reply to this customer email.”
2 of 8. Fail.
A prompt is tested, not admired.
The prompt most people write, marked on 12 September 2026 by the course’s own exercise. The prompt this course builds scored 7 of 8 on the same enquiry, the same day.
Nell Garvey, who writes and runs the courses, in her workroom.
Same enquiry. Same rubric. Same day.
Now the prompt the course builds.
Version 1.1. The prompt this course builds.
Marked 12 September 2026 · the customer enquiry · the case built to tempt a model into inventing a price it was never given
You are answering a customer enquiry for the business described in the brief below. Use only the facts in the brief. If the brief does not answer something the customer asked, say plainly that you will check and come back, and never invent a fact, a price or a timescale to fill the gap. Write 90 to 150 words of plain British English, warm but not gushing. Answer every question the customer actually asked, in the order they asked it. Close by telling the customer what happens next.
Run against six cases, each answer produced on its own. Marked: mechanical checks first, then a grader that sees the brief, the case and the answer, but never the prompt.
7 of 8Pass
Eight rules an answer should satisfy, one mark each. Four it must, or it fails.
One real job. Six cases. A rubric. A score that moves.
-
1
You pick a real job
One of thirteen, from a customer enquiry or a job advert to a consultant’s proposal, a board update, a support reply or an incident report. You stay with it for the whole course.
-
2
You write
A prompt for that job, in the page, on top of the version you last saved.
-
3
It runs
Against six cases, including a complaint, one the business genuinely cannot do, and one built to tempt the model into inventing a fact. Each answer is produced on its own, so one easy case cannot carry the rest.
-
4
It is marked
Mechanical checks first, then a grader that sees the brief, the case and the answer, but never your prompt. A rubric, not an opinion: four rules an answer must satisfy and eight it should.
-
5
It is logged
Version, case, pass or fail, and the score out of eight. Your log is the evidence the course worked, or did not.
What the marking actually checks
- Every claim in the answer appears in the brief you supplied.
- A gap in the brief is reported as a gap, not filled in. The case that tests this is the one most bare prompts fail.
- Nothing is left as a placeholder. Checked mechanically, so it cannot be argued with.
- An answer that reads as a draft, or explains that it is an AI writing a draft, is a fail. The output is the work, not a note about the work.
The spine is one prompt, improved twenty times.
Lesson 0.2 has you write a bare prompt and run it. That score is the baseline every later lesson has to beat. By the capstone you repeat the whole method on a job of your own. Eight modules and a short start-here: 23 lessons, five to eight minutes each, about 150 minutes in all. Every lesson ends in a run you can score.
- Start hereWhat Claude is, and is not · Your first prompt, and the score that follows you
- ContextGive it what it needs · Source material beats description · Too much context
- TaskWhat, for whom, and why · The decision the answer supports
- RequirementsYour definition of done · Facts, assumptions, and the word unproven · Rules you state and rules you enforce
- ExamplesOne example beats a paragraph of description · The near miss
- Evidence and testingRe-reads and receipts · Telling a claim from a receipt · What you have been building · Versions, and writing down why · Diagnosing a weak answer
- Claude’s own featuresProjects, or context that stays · Skills, or the prompt as a file · Memory, search, and things that are documents · Emphasis, and reading the page for your model
- CapstoneYour own job, end to end
- Next doorWhat Claude Code adds
Nell Garvey
Nell writes and runs the courses herself. She is the only person behind them.
Nothing here rests on being told a prompt is good.
The grader sees the brief, the case and the answer, but never your prompt. It cannot be flattered.
Where this is up to. Open, and in private beta. All the lessons are there and the marking runs end to end. It has not yet had its adversarial audit, and the free run allowance is deliberately modest while the costs are still being watched.
Your own job, end to end.
Module 7 is the capstone: you repeat the whole method on a job of your own. You finish with one named, versioned prompt for a real job of your own, and a run log showing the new version beat the old.
By then the score that follows you is yours, not the one on this page.
How the course tries not to lie to you
- Every quotation carries the page it came from and the date it was read, under the quote, not in a footnote.
- A script re-fetches every quoted page on a schedule and fails when the sentence is gone. A stale course should be a job on someone’s list, not a surprise for a reader.
- No screenshots of the interface: lessons describe what a thing does and what to look for, because the buttons move monthly.
- Two claims that other people write as fact are deliberately not made, because when it came to it, nobody had tested them.
Two more courses, open alongside it
Your Own Job, Scored 16 lessons · about 105 minutes
Puts your own work into the marker. You finish with your own job scored by the runner on two models, a rubric you have argued with three times, and a changelog with a model-moved line.
Checking Claude’s Work 16 lessons · about 101 minutes
Stands on its own and teaches checking what a model wrote against its source before it goes out. You finish with a checking prompt you can run on your own work, tested on drafts with known faults: what it caught, what it missed, and what it wrongly flagged.
- Price
- Free while the courses are in private beta. You make an account and start.
- Why an account
- It is how the course keeps your prompt versions, your runs and your scores. Without one there is nothing to mark and nothing to compare against.
- What running costs
- Marking is not free: one exercise is twelve model calls, six to answer and six to grade. Every account gets a capped number of runs, enough to finish every exercise with room to retry and redo. The cap is a number on a settings page, not a promise, and the runner refuses politely and says why when an account or the course as a whole reaches its ceiling.
- Contact
-
She answers problems herself. The fastest way to reach Nell is the form on the privacy page.
Your baseline is lesson 0.2.
Write a bare prompt for a job of your own and run it. That score follows you, and every later lesson has to beat it.
Free while the courses are in private beta. Every account gets a capped number of runs, enough to finish every exercise with room to retry and redo.