The corpus
We wrote up 74 creators across music, architecture, law, science, food, design, technology, film, writing, art and dance. Each creator got three to seven steps, written from named sources and coded into 28 process moves. Each step carries an evidence grade: a record made at the time (132 steps), the creator’s own later account (144), or a historian’s reconstruction (33). Legends were set aside in a separate list of myths.
A second coder, given only the codebook, coded every step blind. Agreement was kappa 0.73 overall and 0.76 to 0.81 on the four jobs below.
No universal move
The most widespread single move, working with a partner, appears in about half of creators. No move is significantly tied to one field either (the lowest corrected q is 0.11), so differences between fields are leads, not findings.
Four jobs that recur
| Job | What it means |
|---|---|
| Feed | Take in real material: places, cases, records, other people’s work. |
| Reframe | Change the question being asked. |
| Import | Bring in something from outside the field’s frame. |
| Vary and select | Make many, test them, keep few. |
Each job shows up in about three quarters of creators. All four together appear in only about a quarter of the short summaries, so a three-to-seven-step account often leaves one out.
What controlled studies add
Everyone in the corpus succeeded, so the corpus cannot say what separates success from failure. Controlled research can, and it backs all four jobs with conditions:
- Conventional work with a small atypical injection beats either extreme (Uzzi et al., 2013, on 17.9 million papers).
- More output brings more hits but not a higher hit rate (Simonton, 1997).
- Working alone first and critiquing together after beats brainstorming as a group (Girotra, Terwiesch and Ulrich, 2010).
How it shapes the workflow
Our process maps onto the four jobs: seeds and real evidence feed it, a person’s tensions reframe, a residency step imports, and the selector varies and selects. It also shows a gap in one framework we study closely. Campo Baeza’s documented process has no import step, which is why we test what happens when one is added.
Limits
The creators were chosen and summarised by a language model from its knowledge of named sources, not re-read page by page; the second coder checks the codebook, not the facts. About one entry in five has a woman as lead or co-lead, and about one in eight is from outside Europe and the Americas.
Sources
- Feathers & Wax, creative-processes corpus, code and data in
research/creative-processes. - Brian Uzzi, Satyam Mukherjee, Michael Stringer and Ben Jones, Atypical combinations and scientific impact, Science, 2013.
- Dean Keith Simonton, Creative productivity: a predictive and explanatory model of career trajectories and landmarks, Psychological Review, 1997.
- Karan Girotra, Christian Terwiesch and Karl T. Ulrich, Idea generation and the quality of the best idea, Management Science, 2010.