From a rough card to work your team can discuss and test.

Not every AI task is a button inside the board. AI Assist supports map and content suggestions in the StoriesOnBoard interface. An MCP-connected agent works from its own environment, reads the live map, and can carry out supported operations. Skills provide reusable instructions for that agent.

This distinction matters when someone wants to reproduce an example. A gap review prompt in an external agent is not a claim that the same workflow appears as a built-in one-click feature. Start with the interface and permissions that match the job.

Begin with a product description and the relevant persona. In AI Assist, work from broad goals toward journey steps and then stories. Select the suggestions that fit your scope; review ordering and granularity before treating the result as the backbone.

For the meal-ordering example, Choose dishes is a step and Check ingredients is a smaller capability beneath it. If a suggestion is really a second user journey, separate it. If it repeats an existing story, merge the intent instead of accumulating cards. The useful output is a clearer map, not the highest possible card count.

The documented AI Assist flow starts in a story card's description editor: press Space and choose Write acceptance criteria. Review the suggested content, ask for refinements, and accept it when it represents the intended behavior. The result becomes part of the card description.

For Check ingredients, the team needs to decide which ingredient information is available and how it is maintained. An AI suggestion must not invent an allergen guarantee. The example below narrows the feature to displaying the restaurant-provided information and explaining when that information is unavailable.

Acceptance criteria templates let a team define the structure it wants AI Assist to use. The documented settings are under AI Assist → Custom AI Functions → Write acceptance criteria. Update the instruction, save it, and inspect the next generated result.

A useful template separates observable behavior, failure cases, and open questions. It can ask for Given/When/Then phrasing without forcing every requirement into an awkward scenario. A template governs the shape of a suggestion; it is not a validator and cannot supply unknown business rules.

For a wider task, ask an external agent to inspect the selected release and explain gaps in the journey. A payment step with only a success story might prompt a question about retrying a failed payment. That is a candidate for discussion, not automatic evidence that a feature must be added.

Define the review boundary: one persona, release, or journey. Ask for a list that separates confirmed gaps, possible duplicates, and unanswered questions. Review it before asking for board changes, then verify the result under the intended steps. Read the skill workflow for a repeatable approach to this review.