Procédure et prompt, ce qui les sépare
Two ways of telling an agent the same thing, which do not live in the same place and do not hold up over time in the same way. **A prompt** is what you retype into the conversation every session. It lives in the exchange, it vanishes with it, and its quality depends on how sharp you are that day. **A procedure** is written once, named, filed, and loaded by the system when it is needed. You do not retype it: you call it. "Run the mail triage procedure" is enough, because the content (which path, which connector, which fallback if the first one drops) is already written elsewhere. The difference is not the length of the text nor its precision, it is **where it lives and who loads it**. A procedure is held by a program, so it can also **forbid**: an agent attempting an action outside the procedure is blocked by the system, not by you watching.
Strengths
- Gives a one-sentence checkable diagnostic: the length of prompt needed measures the architecture debt
- A procedure can forbid an action, a prompt can only advise against it
- The content survives the session closing, hence a tool or model change
Limitations
- Writing procedures before a move is genuinely repeated produces dead structure
- A frozen procedure becomes a constraint when the need moves, it has to be revised like code
Best for
- Deciding what deserves a file rather than a sentence retyped every session
- Explaining to a team why a reliable agent is not steered with better prompts