Chalmers Brothers has spent years teaching Vistage groups that a request isn't complete until both sides agree on what would satisfy it. He calls this the conditions of satisfaction: the specific, often unstated standard that determines whether a promise was actually kept. Most breakdowns in communication, his work argues, come from skipping this step. Two people think they agreed on something, and only later discover they were picturing different outcomes.

Anthropic just published a prompting guide for its newest model tier, Claude Fable 5.1. It's written for developers, full of API code samples, but read past the syntax and it's solving the exact problem Chalmers has been describing for years, just for a "person" who can't infer anything you didn't say. An AI agent doesn't fill in the gaps the way a colleague might. So Anthropic's fixes amount to a checklist for making conditions of satisfaction explicit, because the alternative is a model that executes literally and calls it done.

A few examples.

1. The request has to state what "done" means, or the agent stops short of it. Left alone, the model will sometimes describe the next step instead of taking it, or pause to ask "should I apply this?" for something the original request already covered. Anthropic's fix is a prompt that says plainly: the user isn't watching in real time, so proceed on any reversible action and stop only for genuinely destructive ones. That's a conditions of satisfaction problem. The agent didn't know that "handle this" meant "handle it all the way through," so it asked permission instead of finishing. The fix isn't a smarter model, it's a clearer promise.

2. Silence isn't agreement, and it isn't progress either. By default, the model does more and says less, which means a person can go minutes without knowing whether anything is happening. Anthropic's fix is to specify exactly what a progress update should contain and when to give one. In Chalmers' terms, this is making the commitment visible while it's still being kept, not just at the reveal. If you want to trust delegated work, you need a way to check in on it before it's finished, not just after.

3. Scope has to be a boundary, not a suggestion. The model will sometimes fix nearby code or extend behavior nobody asked for; helpful in intent, but it quietly redefines the promise. Anthropic's fix is to treat the original request as the deliverable's hard edge and route "other things worth doing" to a list for later instead of acting on them now. That's the conditions of satisfaction cutting both ways: it's not just about doing enough, it's about not doing more than was agreed to.

4. What gets carried forward has to be specified, or it gets lost. When a long working session is compressed to save space, the default summary can drop constraints, decisions, or exact numbers, the very details that defined what was promised in the first place. Anthropic's fix is a checklist: preserve what was decided, what's still open, and anything hard to reconstruct later. Every handoff, AI or human, has this same failure mode. The condition of satisfaction from the original conversation doesn't survive translation unless someone insists it has to.

Chalmers has been making the case for years that clear requests and clear promises are the foundation of working relationships, not a soft skill bolted onto them. Anthropic just wrote the technical version of the same lesson, for an assistant that will never once assume it understood you unless you tell it exactly what you meant.