A Claude Code rule that finds where the work will break before you hand it over

Today's file is a rule for Claude Code. Claude says "done", you pass the work on, and the breakage turns up later at the person who got it: a client, your accountant, a colleague. The rule has Claude look for that breakage before the handover: at the end of every finished task and in every plan it adds a "What could go wrong" block.

How it works

Claude imagines that the work has already failed and names the reasons. The psychologist Gary Klein called this a premortem, and the file is named after it. For each reason Claude looks for a remedy it can apply right now. A risk with no remedy does not make the list, and a safeguard that already exists does not count as one. A safe remedy Claude applies before it answers and reports as done, with the result. A risky one, or one that depends on someone else, it names and waits for your decision. If a check can settle a doubt, Claude runs the check first: if it finds nothing, there is no item; if it finds a breakage, Claude fixes it instead of describing it. Anything that needs your consent, it asks about before the work reaches people.

The items are numbered, so you can answer with a number. When there is nothing to say, there is no block at all, not even a heading, and there is none in intermediate steps.

In our test we asked in four languages for a plan to move client bookings from a paper notebook into a Google Sheet. Without the rule the plan had no such block. With it the plan ended in three to five numbered items, such as: photograph the notebook before typing anything, share the sheet only with named people because it holds client phone numbers, pick the day the notebook stops. Where a decision was ours, Claude asked for it: which day, who gets access. To "what is 17*23" Claude answered 391 and added no block.

What to change

The file is a first line and twelve short points: nine on how to build the block and three on how to write it. To get the block only in plans, remove "At the end of every finished task and" from the first line. We added one point ourselves. In the first test we only asked what a small backup script does, and the rule's "fix it instead of describing it" made Claude try to rewrite the script and run it. Now the file says that when only a plan or an answer to a question was asked for, Claude changes nothing and runs nothing, names the remedy and waits. On the same question it then listed three problems with their remedies and touched nothing.

Install

The rule goes at the end of ~/.claude/CLAUDE.md, the file with your instructions for all projects. If you do not have it yet, Claude creates it; the note below has the exact request. That folder holds Claude's own settings, so Claude asks before it edits the file: answer Yes. Then close Claude Code and open it again, because the file is read at startup. The check from the note below: ask for a plan to move client bookings from a paper notebook into a Google Sheet, and the plan ends with a "What could go wrong" block: what Claude already took care of and what waits for your decision.