A user is working with an AI co-creator. The first draft is good enough to invite another. Make it sharper. Warmer. More concise. Less like AI. Try three directions. Combine the first and third. Polish the introduction again.
Every new version leaves less certainty about what should ship. Which direction actually meets the brief? What must not change between drafts? When is the remaining change preference rather than quality? The system won’t say, because it can always produce more.
Generation has no natural ending
Generating variants is easy. Generative systems are excellent at “what else?” Finishing requires “what’s ready now?”
The output of each pass is more possibility, and more possibility drives the next pass. The loop manufactures the reason to keep going and never manufactures the reason to stop.
A product that only generates leaves the user to manage possibilities.
Iteration and convergence are different
Iteration opens possibilities. Convergence closes them.
Early in the work, variation is useful. Later, the product should help compare, select, preserve, check, and hand off.
A converging system knows which stage the work is in.
Sequence map
Convergence stages
Use the stages to move one piece of work toward a clear finish. Record the answer to each question before moving to the next stage.
- Intent: name the item to finish and the finish line: what must be true before it can leave this session.
- Exploration: list the directions that are still viable and the constraint each one must satisfy.
- Selection: name the chosen direction, why it wins, and the alternatives set aside.
- Refinement: name the part to improve and the audience, tone, structure, visual direction, or technical approach that must not change.
- Evaluation: check the chosen version against the brief, constraints, sources, and open questions.
- Handoff: name the next person or system and the context they need to continue.
Chat often treats every stage as another invitation to prompt, which is generous early and exhausting late. The user has to manage the process alone.
Example
Convergence stages: the draft from the opening
The user made several requests to a draft, but the decisions needed to finish it remain unclear. The result is more versions and less certainty about which one should ship.
- Intent: Not known yet: What finished draft is the user trying to produce?
- Exploration: Not known yet: Which of the three requested directions are viable for the draft?
- Selection: Not known yet: Why is the combined first-and-third direction the one to use?
- Refinement: Not known yet: Which decisions must stay fixed while the user makes the first draft sharper, warmer, more concise, less like AI, and polishes the introduction?
- Evaluation: Not known yet: What brief and constraints should the combined draft meet?
- Handoff: Not known yet: What must be true before the draft can leave the session intact?
A useful co-creator preserves settled decisions
It remembers what shouldn’t change.
If the user chose the audience, tone, structure, visual direction, or technical approach, later edits should respect those decisions unless the user reopens them. A request to improve one paragraph shouldn’t rewrite the document. A color adjustment shouldn’t redesign the screen.
A co-pilot that forgets what the user decided turns every pass into a fresh negotiation. Scoped edits, locked decisions, visible diffs, and version history turn exploration into an artifact the user can own.
The product can make “done” visible
The product shouldn’t make the final decision for the user. It can make readiness inspectable. If the product helped make the artifact, it can help define what finished means.
Fill-in card
Readiness check
Make “done” visible as questions the user can check.
- Brief
- does the artifact meet the brief?
- Chosen direction
- does it still match the tone and direction the user chose?
- Open questions
- are they resolved?
- Sources
- are they strong enough?
- Sections
- do they agree?
- Rejected ideas
- are they gone?
- Further changes
- are they now preference rather than quality?
- Next system
- is the output ready for the next system—publish, present, merge, send, or assign?
“You could keep polishing” isn’t the same as “this still fails the brief.” A final pass should reduce launch anxiety, not generate another direction.
Sometimes the most useful response is: “We have enough exploration. Two decisions remain. Choose one, and this is ready.”
Finishing is a product capability
Many AI products append a new invitation after every completed task: “Would you also like me to…?” The pattern turns closure into a sales funnel for more activity.
Useful next moves should come from the state of the job. The next move may not involve AI at all. It may be Review the source, Ask the client, Approve the draft, or Wait for the test to finish. If the work is complete, the interface can confirm, offer export or handoff, and stop offering new work.
When the interface keeps proposing more work, people may stop before they know the work is ready.
A convergence audit
Look at one co-creation flow:
0 of 8 done
