Co Creating With AI

When an author is stuck, don't ask the AI to "finish" the work — ask it to open doors.

Prompt template:
"Generate [10-20] variants for [specific element] in [context],
given these constraints: [genre/tone/world rules/character facts].
Keep each variant to 1-2 sentences."

→ Author picks 2-3 variants that resonate
→ "Develop variant #N further: how does this affect [related element]?"
→ Author rewrites the result in their own voice/style

Example (fiction):

"Propose 12 possible motivations for a clockmaker character who
perceives time differently from others, in a magical realism story
about family and inherited talent. One sentence each."

Core principle: the AI produces raw material (variants); the author produces decisions (choices) and the final text/image/score in their own style. The AI never publishes final work — it only expands the option space.

Progress:
- [ ] 1. Define the creative task with explicit constraints (genre, tone, character/world rules)
- [ ] 2. Generate 10-50 initial variants
- [ ] 3. Collect author feedback: selection, critique, direction
- [ ] 4. Iteratively develop the chosen variants
- [ ] 5. Generate alternative development paths before finalizing
- [ ] 6. Finalize: author selects and manually rewrites in their own style
- [ ] 7. Reflect: log what worked for reuse in future sessions

Step details:

1. Define the task narrowly. "12 motivations for a secondary character" works; "write me a plot" doesn't. Tighter constraints produce more useful generations.
2. Generate in bulk. Ask for a range (10-50), not a single "best" answer. At this stage quantity beats quality — the author recognizes value, not the model.
3. Give specific feedback. "Variant 2, because it reveals the conflict with her daughter" directs the next iteration; "I don't like any of these" does not.
4. Develop depth, not breadth, once a direction is chosen. Ask the model to extend the specific selected variant (scenes, dialogue, consequences) rather than regenerating from scratch.
5. Still ask "how else could this go?" even after picking a direction, to avoid prematurely closing off stronger options.
6. Never publish the draft as-is. The author rewrites in their own voice, rhythm, and style — this is non-negotiable.
7. Log successful prompts and dialogue patterns (Obsidian, Notion) to reuse in future sessions instead of reinventing the process each time.

Example 1 (fiction — good outcome): Input: Writer stuck developing a secondary character — a clockmaker who perceives time differently — in a magical realism story. Process: "Propose 12 motivations for a clockmaker who sees time differently" → author selects 3 → "Develop variant 2: how does his gift affect his relationship with his daughter?" → model produces 5 candidate scenes → author selects and rewrites entirely in their own prose style → AI used afterward only to check timeline/character consistency. Output: Story finished in 3 weeks instead of 2 months; the secondary character became a reader favorite; published in a literary magazine. Author's comment: "The AI didn't write it for me, but it showed me 12 doors I didn't know existed."

Example 2 (visual collaboration — good outcome): Input: Illustrator can't find a visual language for a solarpunk book cover. Process: 20 Midjourney variants generated from an atmosphere/mood description → two directions shortlisted → palette and composition refined through 3 more rounds → final version hand-finished in a graphics editor to match the illustrator's established brand style. Output: Cover instantly recognizable as the illustrator's own work, but with a composition they wouldn't have arrived at alone.

Example 3 (music — good outcome): Input: Composer needs a variation on a theme for the second movement. Process: Suno prompted for 8 harmonic variations with a specified mood → one selected → manually orchestrated and adapted to the instrumentation of the rest of the piece. Output: New theme integrates seamlessly into the score, fully in the composer's established style.

Example 4 (bad outcome — pitfalls ignored): Input: Blogger under deadline asks ChatGPT for "a finished article about remote work productivity," no constraints given. Process: Single draft generated in one shot → lightly reworded for 10 minutes → published as-is, no fact-check, no consistency pass. Output: Article reads generic, cites an outdated statistic the model hallucinated, and has none of the author's usual voice. Readers comment "this doesn't sound like you." Post quietly underperforms and is later taken down. Root cause: no explicit constraints, no variant generation, no manual rewrite, no verification step — every pitfall triggered at once.