Glossary
What is iterative prompting
In short
Iterative prompting is a way of working with a neural network in which you finish the task over several rounds: you look at the answer, refine the wording or ask the model to correct its own draft.
Prompts · iterative prompting
MMOOZZGG.com editorsChecked: 2 min read
For a child
When you make a bunny out of clay, it rarely looks right the first time. You look at it, pull the ears a bit longer, make the tail rounder and look again. A neural network works the same way: the first answer is often not quite what you need. Then you change the request a little and ask again, until you like the result.
In plain words
It is rare to write a task on the first try that a neural network answers perfectly. So people work in rounds. You send a request, read the answer and notice what is missing: too long, the wrong tone, an important detail left out. Then you add a clarification and send it again. Sometimes you do not change the request but ask the model to improve its own answer by your comments. In three rounds, a draft of a greeting for colleagues turns from a dry list into a warm short speech: first you ask to remove the stiff official words, then to add the name of the person celebrating, then to cut it in half.
For business
For a company, this approach cuts the number of redos: in a few short rounds an employee brings a letter, a report or a reply to a client to the accepted form. Save the final wording in a shared library, or everyone will go through the same rounds again. The costs are people’s time and tokens: in a chat, each new round sends the whole conversation to the model again, so long dialogs get more expensive. The risk is that edits fit one good example while answers on other cases get worse. Before you build such a prompt into a work process, run it on a set of real tasks. This check is usually part of the pilot plan.
For a specialist
Iterative prompting (iterative refinement) is a cycle of “generate, evaluate, edit” in which the model’s input or output is improved over several passes. The editor is either a person or an automatic optimizer that changes the wording based on errors on a test set, or the model itself. The Self-Refine method (Madaan and others, 2023) uses one model in three roles: it writes a draft, gives feedback on it (self-feedback) and rewrites the text by that feedback. It repeats the cycle with no extra training, labeling or reinforcement learning. On seven tasks, from dialog replies to math reasoning, the authors got about 20% absolute gain on average over one-shot generation. Anthropic’s guide notes that prompt chains most often follow a self-correction scheme: draft, check against criteria, revise, with each step as a separate API call so it can be logged and evaluated. The cycle stops on a quality metric or a limit on passes, otherwise latency and cost grow.
Example
A marketer asks a neural network to write a product description, gets a text that is too salesy and writes: “remove the exclamation marks, add the size and the material”. The third version goes to the website, and the final request is saved as a template for the whole category.
Related terms
Questions and answers
Sources
Not sure whether your company needs this?
AI solutions map