Reverse Prompting

"Reverse prompting" names two related but distinct ideas. Both invert the default direction of an LM interaction — and conflating them is a common source of confusion, so learn them as separate terms.

Sense 1 — The Flipped Interaction Pattern

Cataloged by White et al. (2023), "A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT" (arXiv 2302.11382), the Flipped Interaction Pattern makes the model ask the questions:

From now on, ask me questions one at a time until you have enough
information to <goal>; then produce <deliverable>.

The instruction has three load-bearing parts — drop any one and the pattern degrades:

Part	Example	Failure if omitted
Goal	"…to design my workout plan"	Model asks aimless questions
Cadence constraint	"one at a time"	Model dumps a 15-question survey in one turn
Termination condition	"until you have enough information… then produce…"	Interview never converges to a deliverable

Why flip? The model usually knows what information the task needs better than the user knows what to volunteer. Flipping moves the elicitation burden onto the party with the checklist — the same logic behind an intake interview at a doctor's office.

Sense 2 — Reverse prompt engineering

Working backward from an output to the prompt that could have produced it: given an artifact (an article, an exam question, a piece of styled copy), reconstruct a prompt that generates artifacts of that kind. Uses:

• Learning prompt craft — dissecting excellent outputs into reusable prompt structure.
• Style transfer — deriving a prompt that captures a voice or format from examples of it.
• Output forensics — inferring what instructions likely produced a given generated text.

This sense powers the exemplar-MCQ workflow in reverse-prompting-patterns and, at pipeline scale, storm-for-exams.

Reverse prompting as pedagogy

Flipped interaction is not just an elicitation trick — it implements active recall (a.k.a. retrieval practice), the best-replicated result in learning science. The testing effect: retrieving knowledge from memory strengthens retention far more than re-reading it. A model that asks you questions forces retrieval; a model that answers your questions permits passive review.

Deployed examples:

• Quiz-me loops — "quiz me on X, one question at a time, grade my answers" (mechanics in reverse-prompting-patterns).
• Socratic tutoring — Khan Academy's Khanmigo and Claude for Education's learning modes are explicitly built to guide with questions rather than hand over answers.
• This site's flashcards — spaced retrieval scheduled by SM-2; see how-this-site-works.

co-storm applies the same inversion to research: its Moderator generates the questions the user didn't know to ask.

Choosing the sense

• Building an app that gathers requirements, preferences, or diagnostic info → Sense 1 (flipped interaction).
• Trying to reproduce or systematize the generation of artifacts you already have exemplars of → Sense 2 (reverse prompt engineering).
• Building a tutor or study tool → Sense 1, aimed at retrieval practice.

Key terms

• Reverse prompting — umbrella term for interaction inversions: either the model asks the user questions (flipped interaction) or the engineer works backward from outputs to prompts (reverse prompt engineering).
• Flipped Interaction Pattern — prompt pattern from White et al. 2023 (arXiv 2302.11382): instruct the model to ask questions one at a time until it can achieve a stated goal, then produce the deliverable.
• Goal / cadence constraint / termination condition — the three load-bearing parts of a flipped-interaction instruction: what to accomplish, how questions are paced, and when to stop asking and produce output.
• Reverse prompt engineering — reconstructing a prompt that could have produced a given output, for prompt-craft learning, style transfer, or output forensics.
• Active recall (retrieval practice) — study technique of retrieving information from memory rather than re-reading it; what quiz-me loops operationalize.
• Testing effect — the empirical finding that being tested on material strengthens long-term retention more than restudying it.
• Socratic tutoring — teaching by asking guided questions instead of stating answers; the design principle behind Khanmigo and Claude for Education learning modes.

See also

• reverse-prompting-patterns — concrete implementations in application code
• co-storm — reverse prompting as a discovery interface
• how-this-site-works — retrieval practice in production