Skip to content
Jonathan Levesque Michaud

Longevity of AI Instructions

Longevity of AI Instructions

· Beginner
Requires
  • ai

Longevity of commands

There are 4 levels of longevity. The shortest one is per message and the longuest/permanent one are account-bound.

Level 1 — In-Message

Scope : Single response only

Think step by step and return only the code

Dies after that one reply. The next message starts fresh with default behavior. Good for one-off overrides.

Level 2 — In-Conversation

Scope : An instruction given early in a conversation that the model tries to carry forward.

For this entire conversation, always respond in JSON
and assume we are using React 18 with TypeScript

Caveats:

  • Degrades over long conversations as context fills up
  • Not guaranteed — the model can drift back to defaults
  • Reinforcing mid-conversation helps (as a reminder, always…)
  • No hard enforcement mechanism

Level 3 — Project-Level

Scope : Every new chat within a project

Powerful option to avoid redundancies when you have a clear objective in mind.

For example, You need to create a new infrastructure which you need to submit to the board of directors. You create two projects : one with high-tech jargons and diagrams, the second that translate text that non-technical coworkers would understand.

Level 4 — Account-Level

Scope : Every conversation across all projects

If you have a specific taste that applies to any chat by default, this would be the right place.

Conclusion

I prefer to use instructions per project that serves a specific purpose. If some exception would be well-suited for a specific response, I use an in-message command.