Steering Notes vs Reminders

~22 min · steering, reminders, noise

Level 0Window Watcher

0 XP0/50 lessons0/13 achievements

0/100 XP to next level100 XP to go0% complete

The two failure modes

Reminding fails by being weak: 'be careful,' 'remember the rule,' 'don't forget X.' The model nods and proceeds. Steering fails by being noisy: long preambles repeated every turn that the model starts to ignore. The cure for one is not the cure for the other.

Steering changes operating state

A steering note has a structure the model can act on: what changed, what is now canonical, what is now forbidden, what is the next safe step. It can be checked against future actions. A reminder cannot.

Compress reminders, expand steering

If you find yourself reminding the model of the same thing every turn, that thing should move to the spine (where it lives once and binds always). If you find yourself steering across many topics, compress to a structured list — bullets and short rules survive long sessions better than paragraphs.

Code

Reminder vs steering, side by side·markdown

REMINDER (weak):
  Please remember to be careful with the database.

STEERING (strong):
  CHANGE: production DB is now read-only for this session.
  FORBIDDEN: any INSERT/UPDATE/DELETE against prod.
  ALLOWED: queries against the staging replica.
  NEXT: confirm you understand by listing forbidden statements.

Steering checklist·yaml

good_steering_note_has:
  - what_changed
  - what_is_now_canonical
  - what_is_now_forbidden
  - next_safe_action
  - explicit_acknowledgment_request

External links

OpenAI compaction guide

Exercise

Take three reminders you have given an AI recently. Convert each into a steering note with what-changed, what-is-now-canonical, what-is-forbidden, and next-action sections.

Hint

If the reminder cannot be converted to steering, the rule was probably never important enough to repeat in the first place.