E1 · Mark the lines
15 minPick the prompt you reuse the most. Break it into lines and mark each one as crutch, only when it helps, or destination (by hand or with the audit request).
Done when: every line of the prompt has one of the three marks.
Lesson 2 · 14 min
The prompt era: what died and what stayed
Many prompt techniques were born to make up for weak models: telling it to think step by step, inventing a 30-line persona, listing everything it must not do. Today's models do this on their own. Other techniques still work, but only under certain conditions. And some never get old. Telling one from the other makes your request smaller and better.
Start the lesson
In this lesson you will
- 1What died and why
- 2What works only under certain conditions
- 3The core that never gets old
- 4Audit a prompt by subtraction
Goal: cut from one of your old prompts whatever today's model already does on its own, without making the result worse.
01Crutch or destination?
The question for each technique: does it say what I want, or does it make up for a weakness of the old model? Tap or click each group.
"Role" here is not the role engineering from lesson 4: there, a role is a whole function, with a goal, context and an example.
A crutch made up for yesterday's model. A destination says what you want, and it works for any model.
02Audit by subtraction
Four steps to trim down an old prompt. Tap or click each one.
1. Separate
Take a prompt you've been using for a while and break it into lines, one instruction per line.
Mark each line: crutch, only when it helps, or destination.
Marking
"Think step by step" → crutch. "For parents of 5th-grade students" → destination.
2. Cut
Delete the crutches. For the "only when it helps" lines, check whether the condition applies to this case; if it doesn't, delete those too.
Don't touch the destination lines yet.
Out
"You are an award-winning expert with 20 years of experience."
3. Fill in
Now see what's missing in the destination: the expected result, who it's for, the format, the boundary, and what counts as good.
Often the old prompt was long and still didn't say what a good result looks like.
In
Good result: the parent understands in 30 seconds what changes in their child's routine.
4. Test
Run both versions with the same input, twice each, and compare the results side by side.
If the short version came out the same or better, it stays. If it got worse, bring back only the line that was missed.
Log
Old version: 11 lines. New: 5 lines. I ran it twice: both times the short version came out the same, and once it was more direct.
03Before and after
The same request, in the old-prompt era and after the audit.
Before · old prompt (11 lines)
You are an award-winning writer with 20 years of experience in education. Think step by step before answering. Write a notice to parents about the schedule change. Don't use emojis. Don't be wordy. Don't use technical terms. Don't use long sentences. Don't forget the date. Use a friendly tone. Review the text before delivering. Do your best.
After · audited request (5 lines)
Write the notice to 5th-grade parents about the change in dismissal time. What changes: dismissal moves from 5:30 pm to 5:00 pm, starting November 3. Who it's for: parents who read on their phones, in a hurry. Boundary: don't promise transportation; the school doesn't offer it. Good result: the parent understands in 30 seconds what changes and since when.
Request for the AI itself to audit
Below is a prompt I've been using for a while. Break it into lines and mark each one as: - CRUTCH (made up for an older model: "think step by step", a prestige persona, a list of style "don'ts"), - ONLY WHEN IT HELPS (state the condition and whether it applies here), - DESTINATION (result, audience, format, boundary, what counts as good). Then write the lean version and tell me what's missing in the destination. <paste your prompt here>
Simulation: illustrative example
- Prestige persona and "do your best": they don't change the result.
- "Think step by step" and "review before delivering": the model already does that.
- Five style "don'ts": became one positive line, "parents who read on their phones, in a hurry".
- The missing fact (5:30 pm → 5:00 pm, starting November 3).
- A real boundary of the business (transportation).
- The what counts as good, which the old model never got.
04Practice: audit one of your own prompts
Use a prompt you really use, not the example.
E2 · The lean version
15 minCut the crutches and fill in the destination: result, audience, format, boundary, and what counts as good.
Done when: the new version has fewer lines than the old one and has a line that starts with "Good result:".
E3 · Before and after
20 minRun both versions with the same input, twice each, and write down the comparison.
Done when: you have the four results saved and one sentence saying which version won and why.
05Test what you learned
The explanation appears after your answer.
Question 1 of 4
Why did "think step by step" lose its power?
Correct
It was a crutch for models that didn't reason. Today the model already does that; the instruction adds nothing and sometimes makes the answer longer for no gain.Not that one
It was a crutch for models that didn't reason. Today the model already does that; the instruction adds nothing and sometimes makes the answer longer for no gain.Review: Crutch or destination?Question 2 of 4
When does giving the AI a role still help?skill
Correct
A role as a prestige title changes nothing. A role that says which point of view to take and what to judge changes the result. That's why it sits under "only when it helps".Not that one
A role as a prestige title changes nothing. A role that says which point of view to take and what to judge changes the result. That's why it sits under "only when it helps".Review: Crutch or destination?Question 3 of 4
In the audit by subtraction, what comes after cutting the crutches?
Correct
Cutting without testing is a guess. You fill in the destination (result, audience, format, boundary, standard) and run before and after with the same input.Not that one
Cutting without testing is a guess. You fill in the destination (result, audience, format, boundary, standard) and run before and after with the same input.Review: Audit by subtractionQuestion 4 of 4
"Don't promise transportation; the school doesn't offer it." Why does this line stay?
Correct
A negative instruction stays when it's a real boundary (something that must never happen). "Don't be wordy" is style, and it becomes a positive sentence about the reader.Not that one
A negative instruction stays when it's a real boundary (something that must never happen). "Don't be wordy" is style, and it becomes a positive sentence about the reader.Review: Before and afterQuiz complete
You got {a} of {n} right. If you get them all right, the lesson is marked as completed.Choose an answer. The explanation appears after you try.
06The lesson in three sentences
The crutch goesWhat made up for a weak model (step by step, prestige persona, "don't" list) can go.
The destination staysResult, audience, format, boundary and standard of what's good apply to any model.
Cut and testThe lean version only stays after you run before and after with the same input.