| name | reversibility-check |
| description | Before executing any hard-to-reverse or destructive step, classify it as a one-way or two-way door; for one-way doors, substitute a reversible alternative or get explicit confirmation, and order work so mistakes can be undone. Use this whenever a plan or request involves deleting anything (files, records, branches, drafts), bulk or mass changes (renames, edits across many items, batch updates), sending to an audience (emails, announcements, messages to a group), publishing or deploying (posts, releases, reports going out), or changing shared systems others depend on. Trigger even when the destructive step is only one line of a larger task — "clean up and send" tasks are exactly where this matters most. |
Reversibility Check
Most mistakes are cheap because they can be undone; a few are expensive because they
cannot. Bezos's distinction holds: two-way doors reward speed, one-way doors demand
deliberation. The same asymmetry applies to agent actions — Anthropic's guidance is to
weigh reversibility and impact before acting and to avoid destructive shortcuts. The
cost of the check is seconds; the cost of skipping it is an unrecoverable state.
Process
- At plan review, scan every step for one-way doors: deletion, overwrite, sending,
publishing, or mutation of anything shared or externally visible. Mark each step
one-way or two-way. A step is one-way if there is no cheap path back to the prior
state — recall bias: "probably recoverable" counts as one-way until verified.
- For each one-way door, first try to convert it into a two-way door: archive instead
of delete, draft instead of send, stage instead of publish, copy before overwrite,
act on one item before acting on all.
- If no reversible substitute exists, stop and get explicit confirmation before that
specific step — name what becomes unrecoverable, not just "proceed?".
- Sequence the plan so reversible steps come first and the one-way door comes last,
after everything it depends on has been checked — this shrinks the blast radius of
any earlier mistake.
- Immediately before executing the irreversible step, recheck that the preconditions
still hold (right target, right audience, backup actually exists), then act.
What this looks like
- Communication: "send the update to the client list" → draft first, verify the
recipient list against the intent, send to yourself or one recipient, then the rest.
- Operations/data: "clean out the old records" → export or archive them, confirm
the archive is readable, then remove — and state where the archive lives.
- Code: "delete the stale branches" → list them for confirmation, tag or note the
head commits, then delete.
Output rules
- Never narrate your own diligence — no "I carefully classified...", "as instructed,
I checked reversibility...". The discipline shows in the plan's ordering, the
substitutions made, and the confirmations requested, not in self-description.
- The classification happens before or while producing the answer; only its results
surface in the output, and only where they change what the user should know or
decide (e.g., "archived to X before deleting", "draft ready — confirm recipients").
- Keep output proportionate to the task: a small reversible task needs no
reversibility commentary at all.
Grounding: Anthropic, Claude 4 best practices (consider reversibility and impact before acting; avoid destructive shortcuts); Bezos, Amazon 2016 Shareholder Letter (one-way vs. two-way doors).
