Teardownspublic

Meta-prompting

Don't write a good prompt — hand over all the raw material and let the AI carve it down. Nowhere in the three steps is the result measured.

The method is right, and nowhere in its three steps is there a place that measures whether the prompt it produced is any good. A sixteen-minute video lays out one person’s meta-prompting style — don’t try to write a good prompt; hand the AI every scrap of raw material and let it do the carving. Half concept, half live demo. What I actually took from it was not the procedure but one sentence about what the question-eliciting step is really for, and what I couldn’t take was any evidence that the method pays off.

three steps — dump everything → carve it down → paste into a fresh window ① dump the context throw all of it in, unrefined paste meeting notes verbatim ② carve the prompt down make it ask back, nail success criteria compress to fit the target tool ③ review the output a human edits with domain knowledge then paste into a fresh window ⇒ nowhere in the three steps is there a place that asks whether the prompt came out good Step ③ is where a human edits, not where anything is measured The fresh-window rule has a clear reason — models do best when the least context is loaded
Three steps, zero gates. The important thing in this diagram is not the boxes but what isn't between them.

What the video says

The definition is one line — have the AI produce the prompt you’ll feed to the AI. Not “build me X” but “I’m going to build X; write me an optimised prompt worth feeding to an AI.”

And three things to get right.

  • Elicit questions. Don’t stop at “write me a prompt”; append “if you need more context to write a good prompt, ask me.”
  • State success criteria. Not “make me an awesome landing page” but “the core value has to be visible in the first mobile viewport, and every button and form has to actually work.”
  • Convert for the execution environment. Each destination tool needs something different — stop conditions for one, constraints for another, framing and lighting and camera for image generation, source standards and verification method for research.

Naming the antipattern was the good part. Going straight in with “I’m building a CRM, write me a meta-prompt” leaves you trapped inside a frame the AI invented on its own. Nothing of yours went in, so nothing of yours comes out, and once a frame sets you only ever edit within it.

One practical tip landed. If you have requests from a meeting, paste them verbatim rather than tidying them up. The moment you refine, only as much as you understood survives.

The technique — why expand first, then cut

Writing 4,000 characters from the start preserves far less detail than expanding fully and then carving down to 4,000. The presenter’s analogy is exact — not being able to write long is the beginner’s problem; the hard part is cutting a long piece down to its points. It’s why school sets minimum word counts and university sets maximums.

And then you start in a fresh window. “Every LLM is at its best early, when the least context is loaded.” There’s no reason to drag along the debris accumulated while building the prompt.

The reason the front end matters more on long jobs is stated too. If you’re going to burn tokens on something that runs for days, starting in the wrong direction costs proportionally more.

what actually got cut in the compression step — the criterion is one line cut — the AI settles these on its own core data model · tech stack · product purpose login · the data-model half of the role system required features → one line each kept — the AI can't guess these roles · expectations for the artifact design — down to background, card and rule tones security · quality level ⇒ one criterion — "can the AI settle this without being told?" ⚠ that criterion cannot be checked afterwards. Cut it and it went well — was it safe to cut, or luck? So the benefit has never been measured — the presenter ends on "I'm not sure this is worth teaching"
The rule for what to cut is clear. There is simply no way to confirm the cut was right.

What broke when I checked

The compression criterion can’t be verified after the fact. “Can the AI decide this dynamically without it being in the prompt?” is a clear rule to apply, but when you cut something and the result comes out fine, nothing distinguishes “safe to cut” from “got lucky.” You’d have to run the same request with and without, and the video contains no such comparison.

And the payoff has never been measured. Nowhere in sixteen minutes is there a figure for whether this beats not doing it, or by how much. The presenter closes by saying “even having done all this, I’m not sure it’s worth teaching” — and that honesty is what raises my confidence in the material. It’s a style share, not a product.

The most valuable sentence is about a side effect, not the procedure. Having asked for questions, no prompt came back — a flood of questions did, and the presenter calls that “already a huge success.” Then only three get answered and the rest are delegated. The line that follows is the core of the video.

Looking at just questions 1, 2 and 3, you can see how thin my original information was.

The real body of the question-eliciting step isn’t a better prompt — it’s showing me where the holes in my own context are. That is a diagnostic instrument, not a prompting trick.

⚠ Though delegating everything makes it pointless. The presenter warns about this: hand it all over and the reason for running the procedure disappears.

Held against my own setup

Steps ① and ② I already had, in code. My pipeline drops the source spec in whole and unrefined, then a deterministic stage cuts it to what’s needed. The shape is the same — except the thing doing the carving is code rather than an LLM, so what got cut and why is reproducible.

Step ③ I hold more strongly. Reviewing the artifact doesn’t end at a human edit; it’s a gate. That is precisely the slot missing from these three steps.

The one thing I didn’t have is single. I have always used “ask me questions” to get better output, never to see what I left out. Same sentence, different purpose, and you read the answer differently.

Verdict

What Call
Question-eliciting as a context-hole diagnostic Adopt. There’s a slot for it before a spec goes in
Expand first, then cut Already doing it. Difference is that code does my carving, so it reproduces
Start in a fresh window Already doing it. The stated reason is worth keeping
Adopt the three-step procedure wholesale Unnecessary. ①② exist and ③ is stronger on my side

The lesson this piece paid for: methodology videos are worth something for their side effects, not their procedures. The three steps were things I was already doing. What survived was one line thrown out in passing — “three questions in, you can see how thin my information was.”

And a method with no place to measure gets neither better nor worse. Whether this pays off is currently unknowable, and finding out means running the same request both ways. The presenter ending on “I’m not sure” is, for that reason, an accurate self-assessment.