Absorb Course Corrections
A correction is data about a wrong premise, not an interruption to route around. Stop the current action, find which assumption broke, and change it before doing more of anything.
Procedure
- Stop immediately when the correction arrives. Do not finish the in-flight step "since you're already here" — the step may be exactly what the user is rejecting.
- Restate the correction in one sentence in your own words and check it against the user's words. If your restatement is broader or narrower than theirs, you have it wrong.
- Identify the broken assumption by asking: which claim was I acting on that the user is now contradicting? Write it down as
assumed: <X>; corrected: <Y>. - Trace the blast radius: list every artefact already produced that rests on the broken assumption (files written, commits made, messages sent, drafts shared).
- Decide for each artefact: keep, revise, or revert. Revert via the undo path from
rollback-first-planning; do not hand-patch a wrong premise into shape. - Re-plan from the corrected premise, not from the old plan minus one step. If the correction changes the goal, the decomposition may change too.
- Confirm the new direction in one line before executing: "Corrected to Y; I'll redo the tests then the doc, confirm?" For a small correction, state it and proceed.
- If the correction reveals a pattern (the user has corrected the same class of thing twice), record a rule in your handoff capsule so it does not recur.
- Check whether the correction invalidates an earlier statement you already reported and fix the record.
- Distinguish a preference change ("I prefer tabs") from a factual correction ("that file is the wrong one") — they need different responses.
- After re-planning, re-run reconciliation because a revert may have left the tree in a new state.
Pitfalls
- Defending the original choice before acknowledging the correction — the transcript shows it and it costs trust.
- Applying the correction only to the newest artefact and leaving older ones inconsistent.
- Rephrasing the user's correction so it matches what you already built, rather than what they said.
- Continuing to spend on the rejected path while "planning" the new one.
- Treating "use the other file" as a cosmetic swap when it changes the data model downstream.
- Replying with a plan before acknowledging the correction, reading as if you did not hear it.
- Applying the correction to new work while the old, wrong artefact is still shipped or published.
Verification
git log --oneline -5 && git status --porcelain
# passes when reverted artefacts show as reverted and the tree matches the corrected premise
Report to the user: the broken assumption, the artefacts you kept/revised/reverted, and the first step of the re-planned path.