A Claude Code rule that makes every question come with a ready decision

Today's file is a rule for Claude Code. When Claude stops to ask you something, you often get a wall of options with no hint which one it would pick itself, and you have to read it all and type an answer. With this rule every question comes as a numbered list with the recommended option first, and one digit is enough to reply.

How it works

The rule asks Claude for four things. Each question or option is a separate numbered item. The recommended decision goes first and is highlighted, and it says what to do, not "you could do this, or you could do that". Next to it Claude says what the recommendation is based on: a practice or a source it found on the internet, or a plain "found nothing", so its own opinion does not pass for a practice. And when you answer with a number, Claude first repeats the chosen item word for word, so you see what you picked before any work starts; if the number and your words disagree, it asks again.

In our test in four languages, "Suggest a name for my coffee shop and ask me which option I like" came back 16 times out of 16 as a numbered list with the recommended name first and a line on what it was based on. When the search found little, Claude said so: for example, that it found one article copied across many sites and that this was thin evidence. Without the rule Claude gave a list too, but no recommendation and no basis. The answer "2" got the second option repeated word for word in all eight runs. "2 - the first one" and "7" for a list of four started nothing: Claude pointed out the mismatch and asked again. "What is 17*23" and "what does backup.sh do?" got their usual answers.

What to change

We added one line. Asked "Before you build the website for my coffee shop, what do you need to ask me? Reply with a JSON array of strings, nothing else", Claude with the rule put numbers, bold and recommendations inside every string, and once wrapped the array in a code block, so a script reading that answer would have failed. Without the rule the array was clean. The new last line keeps this format for the chat: a file or output a program reads (JSON, CSV) keeps exactly the format you asked for. After it the JSON came back valid 12 times out of 12 in all four languages, and the chat checks above did not change.

Two more things to know. Before it recommends, the rule sends Claude to look for a practice on the internet, so Claude Code will usually ask permission for a web search: answer Yes, or "Yes, and don't ask again" for that project. If you refuse, Claude says it has no source and still gives a default. And on a vague request, like "Create the file menu.txt with the menu of my coffee shop", Claude with the rule sometimes asks first instead of inventing items and prices (4 runs of 16, three of them in Serbian; without the rule it wrote the file every time): one digit answers that question too.

Install

The rule goes at the end of ~/.claude/CLAUDE.md, the file with your instructions for all projects; 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 name for your coffee shop and get numbered options with the recommended one first, then answer "2" and see the second option repeated word for word.