A conversation-first artifact puts a conversation with an agent at its centre. The actor arrives with a question, a goal or something loose, and the agent answers, asks, or acts from inside the exchange. Other surfaces, such as a document or a form, appear within the conversation or open beside it.
The patterns run from what the agent can do and how an exchange opens, through carrying each kind of request, checking before consequential acts, and recovering when the exchange breaks, to how it ends or pauses and what the agent keeps from it.
The starter leaves out work the agent runs without the actor watching, which is the delegated agent’s, and the gradual climb of assistance inside other surfaces, which is the assistance escalation sequence’s.
Patterns
Largest first. A starter is where a project's own language begins: leave out what the project does not need, and add what it misses.
- Tool ecology · An artifact designed as one of the many tools, accounts and agents an actor or a group already works with, rather than as the place where all the work happens.
- Information environment · An artifact shaped as a coherent place made of information — with a shared logic of categories, names, relationships, and access — that holds together across every surface, channel, and moment of use.
- Conversation
- Capability & scope
- Opening (Agent)
- Opening (User)
- Inquiry (User)
- Open request
- Inquiry (Agent)
- Prompt scaffolding · Structuring an open prompt into labelled slots, configurable controls, or pre-submission checks when the actor does not yet have their intent in words and the relevant parameters can be anticipated in advance.
- Extended telling
- Inline interface · Reading or acting on a piece of live interface that appears inside the message stream itself, so dialogue and application share a single surface.
- Plan preview
- Inline confirmation · Gathering a second, in-place commitment before a destructive act, without pulling the actor out of context.
- Agent repair
- User repair
- Abort
- Sequence completion
- Closing
- Disengage without closing
- Citation · Claims carry pointers to the source passages behind them, so the actor can see where a statement comes from and check it by following rather than by searching from scratch.
- Expressed uncertainty · Letting the system's confidence vary visibly across its output—through hedged wording, ranges, or marks on doubtful spans—so the actor knows which parts need checking.
- Provenance marking · Attaching a legible mark of what produced a piece of content when it is made, so actors meeting it later know whether it came from a person, a rule, or a model and how to respond to it.
- Memory · Letting the actor inspect, edit, scope, and remove the facts and preferences the system keeps about them across sessions.
- Undo
Not in the library yet
Patterns this shape needs that have no page yet.
- Waiting for generated output · Showing an answer as it arrives and what is still to come, so the actor can read early, stop it, or move on.
- Rating the output · Letting the actor say an answer was good or bad, and telling them where that signal goes and what it changes.
- Branching a conversation · Editing an earlier turn or retrying from it, so the exchange forks and the actor can move between the branches.
- Showing what the agent can see · Making visible the files, history and memory a turn draws on beyond what the actor attached, so the actor can add to it or take things out.
Sequences
The sequences that invoke these patterns, in reading order.
- Making changes · 1 of its 15 patterns
- Working on the same thing · 1 of its 12 patterns
- The encounter · 17 of its 17 patterns
- Delegating work · 1 of its 17 patterns
- Calibrating reliance · 3 of its 10 patterns
- Governing autonomy · 1 of its 9 patterns
- Memory and continuity · 2 of its 9 patterns