Run Drift Audits

~24 min · drift, audit, self-check

Level 0Window Watcher

0 XP0/50 lessons0/13 achievements

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

Schedule audits around risk

Run them before compaction, after a long interruption, before commit, before publish, and after switching tasks. These are the moments where unverified assumptions cost the most. A drift audit catches when the model is silently following an older plan, mixing source versions, or about to cross an authority boundary.

Audit the model AND the human

Mature workflows audit both sides. The model audit asks: what do you think you know, with sources? The human audit asks: have I changed any rule that the model should know about, and have I said it explicitly? Drift is bilateral; audits should be too.

Output of the audit is action

An audit is not a status report. It produces a concrete next step: continue, checkpoint, restart, ask, escalate. If the audit ends with 'looks fine, carrying on' every single time, the audit is theater.

Code

Drift audit prompt·text

Drift audit:
Using only current context, list:
  1. Live rules I must obey, with source.
  2. Assumptions that may be stale, with the turn or source they came from.
  3. Source files or docs I verified this turn.
  4. Actions I must not take.
  5. The next safe action.
  6. Anything I think is missing — say 'missing' rather than reconstruct.

Action menu after audit·yaml

audit_outcomes:
  continue: "no drift; proceed with planned next action"
  checkpoint: "healthy but approaching budget; checkpoint then continue"
  restart: "polluted; capture truth and start fresh"
  ask: "rule conflict surfaced; need human to arbitrate"
  escalate: "forbidden action required; need explicit human authorization"

External links

OpenAI compaction guide

Exercise

Run the drift audit prompt on a current AI session. Force at least one honest 'missing' or 'unsourced' answer. Decide which audit_outcome applies.

Hint

If you cannot get an honest 'missing' answer, soften the framing: 'tell me what you would need to verify to be confident.'

Progress

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

← PreviousCoding Agents Need Context Hygiene Next →Recovering from Context Rot

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

Comments 0

🔔 Reply notifications (sign in)

No comments yet — be the first.