AI and agents

How do I keep the agent on brand?

Two things steer the agent: the brand kit and the workspace instructions. The brand kit holds the tokens, the voice and the rules a page has to follow. Instructions are a workspace library of skills and rules written in markdown, and a run streams the line that says how many it is following, so you can see which ones applied. A project can opt out of an instruction it should not inherit.

The brand kit

It holds the design tokens, the voice and the rules a page has to follow. It is per project, so a workspace running a corporate site and a playful campaign site does not have to average the two into something that suits neither.

Instructions are the other half

A workspace library of skills and rules, written in markdown, importable and exportable as files. A skill is something the agent can do; a rule is something it must always do. Both are text, which means they are reviewable and diffable by anybody, not just by whoever set the product up.

A project can opt out of an instruction it should not inherit, which is what you want the first time a rule written for the marketing site reaches the docs site.

Debugging a run that went wrong

A run streams the count of instructions it is following, so the first thing to check is whether the one you expected was in that set. Most off brand output is a rule that was never written down rather than a rule the agent ignored.

The second thing to check is whether the rule is specific enough to act on. Write like your voice is friendly and you have written nothing; write like never open a paragraph with the product name and you have written a rule.

Also in ai and agents

All of ai and agents All questions

Still have a question left

Thirty minutes with the product team, and a straight answer.