The prompt chaining takes advantage of LLM’s ability to naturally self -can. It improves the precision rate and the clarity of the results while drastically reducing the risk of hallucinations.
All LLM can hallucinate. You have surely already noticed it using Chatgpt, Claude or Gemini … But when you realize it and ask the AI to check if a source or invented information really exists, it corrects itself and admits its wrongs. This is precisely the basis of the prompt chaining, a new technique for prompt language models developed to improve precision, clarity and reduce the risk of hallucinations.
What is the prompt chaining?
The prompt chaining, although already popularized in 2024, remains little used in generative AI projects. Its operation is however simple and robust. The goal is to cut a complex prompt into different individual prompts. The result of each prompt being chained (add) to the following prompt. The objective is to have very simple individual prompts allowing the model to treat only one task at a time. The general precision is drastically increased, and the risk of seeing invented results decreases sharply. It is possible to “chain”, one, two, three or even dozens of prompt in the most complex cases.
To chain a quick, nothing could be simpler. Just identify the different tasks to be accomplished to reach the end result. Once the prompt has been cut into prompt, it is enough to subject them one by one to the model with the previous result of each output. The whole thing takes place automatically in your script with different API calls. Anthropic advises to formulate simple and clear instructions in each prompt and add XML tags around the previous outings of the model to help AI understand the context. The prompt chaining process requires more time in elaboration than using a simple prompt (do not hesitate to iterate several times to find the winning formula) but the precision gain is truly notable.
Very basic example of an effective prompt chaining to achieve daily watch with OPENAI GPT-4O:
Prompt 1 (With GPT-4O + Web search): “Make a current event on (subject) over the last 24 hours. Use only notoriously credible sources. Write a full watch note with a clear and precise language.”
Prompt 2 : “Here is a topical watch on the subject
Prompt 3 : “Here is below, the text written of a day before the last 24 hours about (subject). You will also find advice for improvement, below. Apply the 10 advice for improving the text. Only give the corrected text. Here is the monitoring text to correct:
Why is the prompt chaining so effective?
The prompt chaining is probably one of the most powerful methods to obtain reliable and relevant answers on a multitude of questions. But why is it so effective? Be careful is all you need, as the founding research paper would say. Indeed, language models are based on a attention mechanism allowing them to effectively focus on key elements of a sequence. By cutting out a complex task in several simple steps, the prompt chaining allows the model to fully focus its attention on each sub-probliseme, reducing the cognitive load and promoting structured reasoning, step by step.
In which case should you use the prompt chaining? Research has shown that use cases where several stages (even very simple) are necessary to achieve the result are generally better managed by prompt chaining techniques. For example for writing a text, data analysis under several dimensions or even the creation of a simple agent using tools. Finally, the use of the prompt chaining is even more relevant to the use cases where the traceability of the response is a key element. Identifying what a problem is a problem can be more complex for example with the chain-of-thought. The prompt chaining therefore takes on its full meaning. On the other hand, it is less useful to use the prompt chaining with models of reasoning. The latter being led to reasoning step by step, the efficiency gain will then be minimal.
To go further, the technique of “LLM as a judge”
Form derived from prompt chaining, the technique of “LLM as a judge” allows you to go even further in the relevance of the results. The goal is to generate a first answer with a model A and correct it with a B model.
Example :
Prompt with LLM A : “Generates the biography of Warren Buffett in 100 words.”
Prompt with LLM B : “Does the biography below have non-factual elements? Simply answer yes or no. No other text element.
If, the prompt with LLM B responds “no”, we can send the response to the user, if he answers “yes”, we can either refuse to display the answer or slightly modify the prompt to identify any errors and correct them in the target text.
But why not use the same LLM in this case? The same model naturally tends to validate his own responses, since he will reproduce the same errors or approximations in his judgment as in his initial generation. Indeed, each model has specific “blind spots” depending on their architecture or their training data in particular. The use of a second model, ideally with more parameters, allows you to avoid self-confirmation. Be careful however, even a more efficient evaluator model can potentially hallucinate (much rarer).
