Human–AI creative practice is a repeatable division of labor. The person defines the purpose, constraints, standards, and final decision. The AI retrieves context, proposes alternatives, identifies patterns, and helps test the work. Good collaboration makes those roles visible and makes correction easy.
Start with a working agreement
Microsoft Research's human–AI interaction guidelines emphasize setting expectations, showing contextually relevant information, supporting efficient correction, and learning from user behavior without becoming unpredictable. NIST similarly stresses that human roles and oversight responsibilities should be explicitly differentiated.
Translate that into a five-line agreement before the first prompt:
- Intent: the change this work should create for a specific person.
- AI role: researcher, divergent ideator, critic, organizer, editor, or simulator.
- Human role: source selection, taste, domain judgment, risk acceptance, and approval.
- Evidence boundary: which source pack is authoritative and which claims require verification.
- Stop condition: what “useful enough to test” means.
Use a four-turn creative loop
- Frame. Give the AI the audience, desired change, constraints, source boundary, and what must not be invented.
- Diverge. Ask for meaningfully different approaches, including one that challenges your preferred direction.
- Interrogate. Require assumptions, weak evidence, likely failure modes, and what a skeptical user would object to.
- Commit. The human chooses, edits, and records why the final direction won.
This loop turns prompting into a traceable creative process. It also prevents a common failure: accepting a polished first response because it sounds finished.
Correctability is a design feature
A system that remembers should make correction more visible, not less. Show which memory or source influenced a suggestion. Let the person reject a connection without fighting the interface. Preserve the original draft when a summary is generated. Record whether a change came from the person, the model, or an automated rule.
These details matter because context can disappear when complex human goals are reduced to measurable signals. NIST warns that representing human phenomena mathematically can remove necessary context. The answer is not to avoid measurement; it is to keep measurements scoped, reviewable, and subordinate to the actual human goal.
What to measure
Do not measure collaboration by output volume. Track whether the practice improves decisions:
- How often does the human substantially revise the AI's first direction?
- Which sources survive into the final artifact?
- How many unsupported claims are caught before publication?
- Does remembered context reduce repeated explanation without introducing stale assumptions?
- Can the person explain why the final artifact has its current shape?
A compact ritual for real work
Use the Idea Seed to Artifact Builder to create the working agreement and proof question. If the project spans multiple sessions, use the AI Memory Planner to decide which raw episodes, decisions, and preferences should carry forward.