A seven-step process for making AI write in any human style

Executive overview

Most AI writing outputs sound robotic because users tweak prompts randomly rather than following a structured extraction process. The fix is building two dedicated AIs: one that extracts a writer's style, one that mimics it.

Pick one style to mimic — your own or someone else's. Combining multiple styles confuses the model.

Train a style extractor and a style mimicker as separate AIs, fed by research-driven system prompts.

The seven-step process

1. Choose a single writing style to mimic (your own or one other author — not a blend)
2. Use an AI search tool to research what specialists look for when mimicking writing: linguistic elements, semantic structure, stylistic patterns, etc.
3. Prompt an AI to write a system prompt for a style extractor, embedding the research findings and asking it to follow current prompting best practices
4. Condense that system prompt using the provider's native prompt generator (Claude or OpenAI) — target two pages, not five
5. Feed four to six diverse writing samples into the extractor to produce a documented style profile
6. Use the same approach to build a second system prompt for a mimicker AI, giving it the extracted style profile and instructing it to treat samples as inspiration, not content to copy
7. Deploy both AIs as custom projects (Claude Projects, Custom GPTs, Gemini Gems) with the relevant system prompt and knowledge base

Choosing a writing model

• Claude Sonnet 4: best general-purpose writing model; articulate on complex topics; tends toward wordiness
• o3: concise and focused; citation formatting is limited (end-of-sentence only, not inline)
• GPT-4.5: capable but slow; likely to be deprecated from the API
• Grok 3/4: stronger for creative or fiction writing; fewer topic restrictions
• Gemini 2.5 Pro: not strong for writing; best reserved for large-document analysis with oversized context windows

Advanced: aggregating outputs from multiple models

• Same-model aggregation fails — a model grading itself picks itself ~95% of the time
• Use a third-party model as the aggregator (e.g., use Gemini to choose between Sonnet 4 and o3 outputs)
• Randomise input order across iterations to reduce positional bias
• If bias persists, force a 50/50 split through the prompt
• Do not let the aggregator rewrite anything — instruct it to select sections verbatim, then use code to assemble the final output
• Allowing rewriting causes drift from source material and degrades accuracy