Skip to content
Mastering Claude

Home / Writing a request that works

Writing a request that works6 minApplication

Making a working request last and reusing it

A request that works is refined in small touches rather than started over from scratch, and then fixed into a reusable template where only the part that changes from one task to the next gets replaced.

A request that produces a good result deserves to be kept and adjusted, not rewritten from scratch on every new attempt. Starting from a blank page each time loses what was already working and delays the moment the request becomes reliable.

Adjusting in small touches

When a response drifts slightly from what was expected, the effective reaction is to change a single part of the request, the instruction, the example or the format, then replay the same request on the same case. Comparing the two responses shows precisely what that change produced. Rewriting the whole request on every attempt mixes several changes at once, and makes it impossible to know which one fixed, or broke, the result.

Placing a long document right at the top

When a request comes with a long document, a contract, a report, a long list, its position in the text changes the quality of the response. According to Anthropic's documentation on long context, dated 2 September 2026, placing long documents and data at the top of the request, before the question, the instructions and the examples, improves performance across the board on all models. This simple reflex, pasting the document right at the start rather than at the end, applies just as well to a contract pasted once as to several files delimited by clear tags.

Fixing a request into a reusable template

Here is what a request fixed into a template looks like. Everything is fixed, except what sits between double braces:

Tu prépares un compte rendu pour le pôle {{POLE}}.
Résume le texte ci-dessous en cinq points, un par ligne, sans introduction.
Garde les montants et les dates tels quels.

Texte à résumer :
{{TEXTE}}

For the next task, two replacements are enough, and the four lines of instruction no longer need renegotiating.

Once a request has been adjusted until it regularly produces a good result, fixing it into a template avoids rewriting it for every similar task. The technique consists of replacing the part that changes from one task to the next, a client's name, a date, a text to process, with a marker written between double braces, and keeping the rest of the request fixed. The same template then serves for twenty different tasks, all that is needed is to replace what sits between the braces before each send. The same reflex for precision applies to a summary, by naming the expected length and angle rather than letting Claude choose, and to a request in several languages, by supplying your own list of terms you want translated one particular way rather than another.

Figure 1

From a first attempt to a fixed template

01
First attempt
An initial request sent on a real case, with a response that is still imperfect.
02
One single change
An instruction, an example or the format is changed, never several at once.
03
Same case replayed
The changed request runs on the same case as the first attempt, to compare the two responses.
04
Version kept
The change that improved the response is kept, the one that made it worse is undone.
05
Fixed template
The stabilised request is saved with markers between double braces in place of what changes from one task to the next.
The diagram follows a request from its first, imperfect attempt through to its form fixed into a template, passing through changes tested one at a time on the same case.
Calibrate it yourself

An executive assistant sends Claude a fixed template request, where the client's name and the meeting date are marked by markers between double braces, to prepare meeting minutes. She replaces these two markers and sends the request for twelve different meetings over the course of the month, getting minutes in the same format each time.

Write, in one sentence, what this situation establishes, and in one sentence what it does not establish.

What to remember
  • Changing a single part of a request at a time, then replaying the same case, shows precisely what that change produced.
  • Placing a long document right at the top of the request, before the question and the instructions, improves performance according to Anthropic's documentation.
  • A marker between double braces flags the spot that changes from one task to the next in a fixed template.
  • Naming the expected length and angle of a summary avoids leaving Claude to choose these two settings alone.
  • The same reflex for precision applies to a request in several languages, by supplying your own list of expected terms.
Do this now

Take a request you send several times a week in almost identical form. Replace the part that changes each time, a name, a date, a text, with a marker between double braces such as {{DOCUMENT}}, save this version in a text file you can reopen, and use it today on your next send by replacing only that marker.

Check the source

Every datable claim in this lesson links here to the public text behind it. A source that does not open proves nothing.