$5. That's what criminology professor Andrew Wheeler spent on AI to co-write his latest book, LLMs for Mortals. His previous book took a year. This one took two months, with roughly half the content generated by Claude Sonnet 4.1.
The workflow is more practical than you'd expect. Wheeler writes in plain text (Markdown and Quarto files), drops prior writing samples into Claude's context window so the model learns his voice, then feeds it detailed outlines for each section. The AI drafts content that matches his style closely enough that Pangram's AI detection tools flagged none of the published passages as machine-generated.
His setup boils down to three things:
- Prior writing samples in the project folder so Claude can mimic his tone
- Custom instruction files specifying preferences (no emojis, no excessive bullet lists, no padding)
- Section-by-section generation rather than asking for entire chapters at once
The cost breakdown is striking. Each blog post runs about $0.50 in API costs. Per-chapter editing still takes 20-30 hours of expert review, which is where the real time investment lives. Wheeler is clear that subject-matter expertise is non-negotiable for catching errors. The AI writes fast but doesn't know what it doesn't know.
One specific tip for academic writers: compile your cited references in BibTeX format before generating text. This prevents the model from hallucinating citations, a well-known problem when LLMs write research-adjacent content.
This is a useful case study in what "AI-assisted writing" actually looks like in practice. Not push-button automation, but a workflow where the human provides structure, expertise, and quality control while the AI handles first-draft generation at near-zero marginal cost.