The brief

4.09[BONUS]

Client checks, the mistakes it can actually catch

Ask it whether its own work is good and it will tell you it is.

Anthropic, who make Claude, found that agents asked to judge their own work tend to praise it, even when the quality is obviously mediocre to a person reading it.

A question with a yes or no answer is different, because there is nothing to argue about, so it can check its own work against it as many times as it takes.

A judgement

Write the client update

you

Ask whether it is professional

Yes, it reads well

you

Send it, with last week's client still named in it

It marks its own work, and passes it.

A yes or no

Write the client update

you

Ask whether any other client is named in it

Yes, the third paragraph names another client

you

Fix it before it goes

It looks, and catches it.

So write the mistake you would least want a client to see as a yes or no question.

“Make sure it is professional” is a judgement, while “no other client’s name appears in this” can actually be checked.

A few of these sit in the top brief and run on anything going to a client, and a check that only matters for one job goes in that job’s skill, which stage six builds.

Paste this into a chat with your ecosystem folder open.

It asks which mistakes would cost you, writes the checks into your standing rules and runs them on the last thing it wrote for a client.

Whether the work is any good stays your call, because that judgement is what your clients pay you for.

Add client checks to my brief

Open the CLAUDE.md at the top of this folder. I want a few client checks in my standing rules, questions with a yes or no answer that you run on anything going to a client before you show it to me. Interview me one question at a time about the mistakes I would least want a client to see. Then write three to six checks, each one a yes or no, never a matter of taste like "is this good". Add them to my standing rules in the same wording and tone as the rest of the file, with a line saying that when one fails you tell me which and why, rather than quietly changing a name, a figure or a date. Show me the change. Then run the checks on the last thing you wrote for a client, or ask me to paste one in.

The finding is from Anthropic’s engineering team, Harness design for long-running application development, March 2026, which notes the problem is worst on subjective work, where there is no yes or no check to run.