Home / Writing a request that works

Composing a complex request without inventing

A complex request is composed piece by piece, who Claude should be, the context, the data to process, what is being asked, an example, the expected format, then a final line that allows Claude to say I do not know rather than invent an answer.

A complex request mixes several types of content, Claude's expected role, the task's context, the data to process, the instruction, an example, the wanted format. Thrown into a single paragraph with no order, these elements force Claude to guess which is which, and one of them, the permission to say it does not know, is almost always the first to disappear.

Putting each piece in its place

Assembling a complex request piece by piece, in a fixed order, removes this sorting work. One practical way to proceed is to arrange, in this order, who Claude should be, the context of the task, the data to process clearly delimited, what is being asked worded as a positive instruction, an input to output example, then the exact format of the response:

Role : vous etes assistant de gestion administrative.
Contexte : ce resume sert a une decision de comite demain.
Donnee : <dossier> ... </dossier>
Instruction : listez les points bloquants du dossier.
Exemple : dossier A -> trois points bloquants numerotes.
Format : liste numerotee, cinq lignes maximum.

Each piece stays isolated from the others. Changing the example or the format later touches nothing else in the request, which makes the next version easier to correct than if everything sat in a single block of text.

Adding a fallback output before it is missed

The last piece matters as much as the others, a safeguard, a line that explicitly allows Claude to answer I do not know rather than invent information absent from the data supplied. Anthropic's documentation on reducing hallucinations puts it this way: giving Claude explicit permission to admit uncertainty is a simple technique that can sharply reduce false information. Without this line written into the request, Claude picks the most likely answer given the rest of the text, even when that answer is wrong.

The lesson on input to output examples finds its natural place here, right before the expected format, to show Claude what a correct response looks like before giving it the output that covers the case where information is missing.