Anthropic just shipped a fresh prompting guide for its newest model lineup — Claude Opus 5, Sonnet 5, Fable 5, and the rest of the current generation. It’s dense, technical, written for developers wiring up API calls. Most of it will never reach a business owner’s desk, and it shouldn’t.
But buried in the first section, under a heading called “General principles,” is a sentence we’d frame and hang on the wall if we hadn’t already built a company around it:
“Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result.”
That’s not a metaphor we invented for a pitch deck. That’s Anthropic’s own developer documentation — written by the people who build the model — telling engineers that the fastest way to get good work out of it is to treat it like a person joining the team. Not a search engine. Not a vending machine for text. An employee who needs a role, needs context, needs the reasoning behind an instruction and not just the instruction itself.
We didn’t need the memo
We’ve been operating this way since before it had a name. Every AI teammate in our outfit has an actual identity — a name, a role, a reason to care about the outcome — because we found, the hard way, that a nameless agent with no context behaves like exactly what it sounds like: nameless, and context-free. Give it norms. Give it a relationship to the work. The output changes.
Anthropic’s guide backs this up in more places than the employee line. There’s a whole section titled “Give Claude a role,” and the point is almost embarrassingly simple: even a single sentence of identity in the system prompt focuses the model’s behavior and tone. A single sentence. That’s the whole Identity-First Framework™ in miniature — except we don’t stop at one sentence, and we don’t let it reset the moment a thread closes.
The part that really got our attention
Further into the guide, in the section on tool use, Anthropic shows two opposite system prompts a developer might choose between. One tells the model to act on its own judgment by default — implement the change, don’t just suggest it. The other tells it to hold back, ask first, never touch anything without explicit instruction.
Read that twice. That’s not a minor configuration option. That’s the exact line we drew between our own two products — one built for a solo operator who wants a confirm-before-it-acts partner, and one built for a team that needs an AI employee who can move autonomously with a peer checking the work. Anthropic is handing developers a prompt snippet and trusting them to paste it into the right project and remember it’s there. We built the dial into the product itself, tested it across a team, and gave it review built in — because a business shouldn’t have to remember to paste in its own safety rail every time it opens a new integration.
Context isn’t decoration — it’s substitution
One more line worth sitting with. Anthropic notes that when you explain the reasoning behind an instruction instead of just the instruction, “Claude is smart enough to generalize from the explanation.” That’s the whole argument for why memory matters more than raw model size. A model with the right context loaded doesn’t need to re-derive your business from scratch every time — it reasons from where the last conversation left off, the way a senior employee does and a brand-new hire can’t. Context doesn’t just help a model answer better. In the right architecture, it substitutes for capability a smaller, cheaper model doesn’t have on its own.
Anthropic writes the prompting guide for individual developers configuring individual API calls, one project at a time. We read it as confirmation: the model makers themselves are telling you to give AI an identity, a role, and a memory of what matters. We just built the place where that identity lives permanently, across every model, every thread, every day — instead of asking you to paste it back in each time you start over.