Personalizing AI writing
The Landscape of Taste Transfer
Selected approaches to personalizing AI writing, with evaluation questions and primary sources.
Selected approaches
The methods below act on different parts of a system. Compare them against a specific writing task before combining them.
| Method | What changes | What must be evaluated | Primary source |
|---|---|---|---|
| Prompt optimization with GEPA | Proposes and evaluates prompt changes using feedback. | Improvement against a held-out baseline; preserve instruction following. | DSPy GEPA documentation |
| Context playbooks with ACE | Revises context through generation, reflection and curation. | Whether the playbook preserves the author’s choices on unseen examples. | ACE paper |
| Model activation steering | Changes selected model activations during inference. | Personal writing-style transfer requires its own evaluation. | Persona-vector research |
ACE’s application to personal writing style remains a proposed experiment. Anthropic’s persona-vector research studies model character traits. It does not establish reliable transfer of an individual writer’s voice.
This report does not establish a percentage of personal taste that any method captures. Editorial judgment should be evaluated separately from surface style.
Evaluation questions
Start with examples you have permission to use and define what a successful result preserves. Compare prompting with retrieval before adding training or model-internal changes.
Define what the writing should preserve
Choose examples that show the author’s editorial decisions as well as their phrasing. Identify the intended audience and what the piece should leave out. Keep some examples out of development so the evaluation can test unfamiliar material.
Compare with a held-out baseline
GEPA uses feedback to propose and evaluate prompt changes, as described in the linked documentation. Compare the result with a held-out baseline. Check whether the revised prompt still follows the task’s instructions.
Use the same writing brief for each candidate. Review factual accuracy and editorial choices separately from whether the prose sounds familiar.
Test combinations as a hypothesis
Combining these techniques is a design hypothesis. A useful evaluation would test whether the combination preserves the author’s choices.
Watch for insufficient examples and interference between changes. A method that makes one passage sound closer to a reference may still fail the task; check the complete output before adopting it.