Speech Act, Not Magic Phrase

Borrow a frame from philosophy of language

Speech-act theory (Austin, Searle) distinguishes what a sentence says from what it does. "Can you pass the salt?" is grammatically a question; functionally it's a request. The same is true of prompts: the surface words matter less than the speech act they perform.

Three useful speech acts for prompting

• Directive — instructs the model to do something. "Summarize." "Translate." "Decide."
• Constitutive — declares what is true in the conversation. "You are a contract analyst." "The user has admin rights." Constitutive prompts shape every following turn.
• Commissive — commits the system to a behavior. "You will refuse to discuss medical dosages." "You will respond in JSON only."

Why this is useful

When a prompt fails, ask which speech act is broken. "It ignored my instructions" — was the instruction a clear directive, or was it buried inside a constitutive claim? "It refuses safe requests" — is a commissive too tight? Naming the act sharpens the diagnosis.

The model's default

Frontier models default to interpreting most prompts as directives unless a constitutive layer (system prompt) says otherwise. This is why an empty system prompt produces a chatty, eager-to-please assistant — the constitutive frame is missing.