Re-Anchoring at Each Turn

~12 min · conversation, anchoring

Level 0Apprentice

0 XP0/100 lessons0/14 achievements

0/120 XP to next level120 XP to go0% complete

What you re-inject every turn

Some constraints are too important to dilute. For these, re-inject on every turn — not via system prompt (which is fixed), but as a short reminder near the top or bottom of the active context.

Candidates for re-anchoring

Refusal policies ("do not give medical dosages").
Output schema reminder for structured-output calls.
Active task name and current step (for agent loops).
Tone constraints that drift over time.

How to inject

A turn-level developer-role message (where supported) or a structured user message marked as system reminder. Make it visually distinct in logs so you can audit it.

Code

Per-turn anchor·python

anchored_messages = [
    *history,
    {"role": "user", "content": user_input},
    {"role": "user", "content": (
        "[reminder] You are still in 'underwriting' mode. "
        "Output JSON matching the verdict schema. "
        "Do not produce free text outside JSON."
    )}
]

External links

Anthropic — Reminders pattern

Exercise

Identify one constraint that drifted in a long conversation. Add a per-turn anchor for it. Re-run the conversation; observe whether drift is now bounded.

Progress

Progress is local-only — sign in to sync across devices.

← PreviousLong Conversations — Compaction Strategies Next →Stateful Tools in a Stateless Model

Spotted a bug or have feedback on this page?Report an Issue

Comments 0

🔔 Reply notifications (sign in)

No comments yet — be the first.