Home / Techniques that change the result
Chaining prompts
Breaking a complex task into a chain of short prompts, where the output of one becomes the input of the next, makes each link verifiable on its own instead of judging a single opaque block.
A complex task handed to Claude in a single message produces a single block of text, hard to correct precisely when something in it has gone wrong. Breaking that same task into several short messages, where the answer to one serves as raw material for the next, changes what can be checked: each link becomes a point where you can read, correct, or stop the chain before continuing.
The most common pattern: draft, review, correction
The simplest pattern fits in three messages. The first asks for a draft. The second has that draft reviewed against named criteria, listing what needs to be checked. The third asks for a corrected version built from that review. Each message is a separate send: nothing stops you from reading the intermediate output, editing it by hand, or rephrasing the next request before sending it.
Message 1 : rédigez un brouillon.
Message 2 : relisez ce brouillon selon trois critères précis, listez les problèmes trouvés.
Message 3 : corrigez le brouillon à partir de cette liste.The third message does not start from scratch: it builds on the list produced by the second, which itself builds on the text produced by the first. This dependency is what makes each link verifiable on its own. A problem in the final result can be traced to a specific link, the draft, the review or the correction, rather than to a single block where nothing distinguishes their origin.
What chaining does not provide on its own
Claude already handles a good part of the reasoning internally in several steps on a complex task, with no manual breakdown needed for that. Explicit chaining across several messages keeps its usefulness for a specific purpose: forcing a stopping point where you look at what has been produced before continuing, or imposing a precise structure you want to follow step by step. It is not the default reflex for every task, it is a tool for tasks where you want to read or correct before the next step goes ahead.
The link with a check's positive control is direct: a check that judges a single opaque block at the end of a chain cannot say at which link the error appeared. A check placed at each link, on the other hand, isolates the exact spot where the chain went wrong.
Where to place the verification point
The most useful verification point sits between two links, before the output of one becomes the input of the next. Reviewing the draft before sending it off for review, or reviewing the list of problems before requesting the correction, costs a few seconds and stops an error in the first link from propagating, apparently fixed only in appearance, into the next two links.
Three links, each verifiable on its own
A person hands Claude, in a single message, the drafting of a paragraph aimed at customers. She then sends a second message asking Claude to review this paragraph against three criteria she names, then a third message asking for a corrected version built on that review.
Write in one sentence what this situation establishes, and in one sentence what it does not establish.
What this establishes: That the corrected version in the third message builds on the list of problems produced by the second message, itself drawn from the text of the first.
What this does not establish: That breaking this task into three messages produced a better paragraph than a single message would have produced.
The three most common miscalibrations
- Too broad This situation shows that breaking a task into several messages always produces a better result than a single message.
- Too narrow This breakdown into three messages proves nothing since it only concerns a single paragraph aimed at customers.
- Beside the point This situation shows that three separate messages take longer to write than a single detailed message.
- A task broken into separate messages makes each step readable and correctable before the next one goes ahead, unlike a single block of text produced all at once.
- The simplest pattern chains three messages, a draft, a review against named criteria, then a correction built on that review.
- Claude already handles a good part of the reasoning alone in several steps on a complex task, manual chaining keeps its usefulness for forcing a precise stopping point.
- The most useful verification point sits between two links, before the output of one becomes the input of the next, not only at the very end of the chain.
Take a task you usually hand to Claude in a single message. Break it into three separate messages, a draft, a review against criteria you name, a correction built on that review. Send the three messages one after another in the same conversation and compare the result you get to what a single message would probably have produced.
These points depend on an interface or a rule that may have changed since this was written. Check them on your own screen before relying on them.
- Before breaking a task down by hand, let Claude handle it in a single message and look on your screen whether its answer already reveals reasoning done in several verifiable steps: manual chaining only adds something if it does not do this on its own.
Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.
- Anthropic, Prompting best practices, section Chain complex prompts consultée le 2026-09-02