Home / Techniques that change the result

Composing a system prompt from several sources

Merging the best rules from several system prompts works like an ordinary piece of writing, provided each imported rule is rewritten for its own context rather than copied as is, and the rules already in force settle any conflict.

A system prompt, or custom instructions depending on the interface used, is the text that frames Claude's behaviour before the first question is even asked, a role, a register, rules to follow under all circumstances. A system prompt found elsewhere, a template shared by someone else or a list of rules published as an example, almost always contains one or two useful sentences amid a lot of detail that does not concern your own use. Copying the whole thing as is also imports what does not apply to your context. Taking a single rule and rewriting it in your own words, for your own situation, is an ordinary piece of writing, the same kind used to draft any instruction meant for Claude.

Rewrite, do not copy

A rule copied word for word often carries vocabulary, examples or constraints specific to the situation of its original author. A rule aimed at a certain type of answer for a given professional audience does not carry the same meaning for a non profit audience. Rewriting the rule forces you to ask what it actually targets, before rephrasing it for the context where it will apply. This rephrasing work is what distinguishes borrowing an idea from copying a text.

Règle trouvée ailleurs : ne jamais utiliser de jargon technique sans le définir.
Règle réécrite pour son contexte : chaque terme métier du secteur associatif, adhérent,
cotisation, assemblée générale, se définit en une phrase à sa première apparition.

Who settles a conflict

An imported rule can contradict a rule already in place. Priority goes to the rule of the system prompt's owner, not to the imported rule: it is the owner who knows their own context and the reasons why their rule exists. Before adding a rephrased rule, you must therefore read it against the rules already in force and check that neither contradicts the other.

Checking that a rephrased rule still triggers the expected behaviour follows the same logic as testing a check before trusting it: a rule never put to the test on a real case has not proven that it works in its new context. Asking a question that should trigger the rule and reading the answer obtained is the simplest way to check it.