Chapter 20 Key Takeaways: Writing and Editing with AI

  • Writing is a multi-stage workflow, not a single act. The seven stages — ideation, outlining, first draft, developmental edit, line edit, proofreading, and audience adaptation — have different AI fit levels. Matching AI use to the right stage is the core skill of AI-assisted writing.

  • AI fit is highest at the edges of the writing process. Ideation (generative, low-stakes) and proofreading (mechanical, well-defined) are the highest-fit AI stages. Line editing — prose quality at the sentence level — carries the greatest risk of voice bleed.

  • AI default style is recognizable and must be actively managed. AI prose is statistically smooth — balanced sentences, careful transitions, calibrated tone, and an absence of the idiosyncrasies that make individual writing distinctive. Without active voice management, AI-assisted work gradually converges toward this default.

  • Style extraction is a reusable asset. Submitting your best writing to AI for style analysis, then saving the resulting style guide as a prompt component, is one of the highest-leverage investments in voice preservation you can make.

  • Voice injection works by volume. The more of your own prose you include in the context window before an AI writing task, the more AI output will sound like you. More of your prose equals less of AI's default style.

  • The distinction between "AI as ghostwriter" and "AI as editor" is fundamental. As ghostwriter, AI brings its voice because you have not provided one. As editor, AI works within your existing prose. Prefer the editor mode for voice-sensitive documents.

  • Expand-then-contract produces better results than single-draft generation. Generating more than you need, then selecting and editing down, is reliably superior to asking AI to write exactly to a target length. AI is a quarry, not a finished product.

  • Request three genuinely different drafts and synthesize. The multiple drafts technique forces broader exploration of the solution space and gives you real alternatives rather than accepting AI's first path as the best one.

  • Name the editing mode explicitly. Developmental, line, proofreading, and tone check edits require completely different prompts. Undifferentiated "edit this" requests produce inconsistent, often wrong-headed outputs.

  • Long-form AI collaboration requires deliberate coherence management. Include the full outline, your introduction, and previously written sections in every generation prompt. Catch voice drift and inconsistencies immediately — they are much harder to fix after assembly.

  • Fact-checking is non-negotiable and non-conditional. Every factual claim in AI-assisted writing must be verified against a primary source before publication. This rule has no exceptions for document type, audience, or time pressure.

  • Reading aloud is the most reliable voice-check method. Voice anomalies that survive silent reading are almost always caught when spoken. Build a read-aloud step into every AI-assisted writing workflow.

  • AI padding is the most common quality degrader. AI reaches target length through redundancy — repeated points, empty qualifications, transitional summaries. Cut aggressively; treat the AI draft as overlong by 20-30% until proven otherwise.

  • The introduction should almost always be written by the human author. AI introductions tend toward the generic. The introduction establishes voice and frame for everything that follows — it is too important to outsource.

  • Audience adaptation is a high-value, low-risk AI use case. Transforming existing content for a different audience (executive brief from full report; LinkedIn from blog post) is systematic work that AI handles reliably. Define the target audience explicitly and evaluate the output carefully.

  • Technical documentation benefits from AI because the genre prioritizes clarity over voice. AI can generate comprehensive docstrings, READMEs, and API documentation reliably — but only when given accurate technical inputs. Technical accuracy requires human verification.

  • Consulting and advisory writing carries specific voice risks. Generic consulting language — hedged recommendations, vague options, liability-conscious softening — is exactly what AI defaults to in professional report contexts. Voice checks and fresh-eye review are essential in this genre.

  • The fresh-eye rule is a structural safeguard. The author who has spent hours on a document cannot read it the way a client will. Temporal distance (overnight minimum when possible) or the read-aloud technique are the mechanisms for recovering fresh-eye perspective.

  • AI writing assistance improves lower-skilled writers more than expert writers. For experts, poorly managed AI assistance can introduce quality regressions. Expert writers need more active voice management, not less.

  • Disclosure norms are context-dependent and evolving. The professional who publishes AI-assisted work is responsible for its accuracy and appropriateness regardless of disclosure. Disclose when the audience has a reasonable expectation to know, or when organizational or professional policy requires it.

  • Workflow documentation is insurance against under-pressure shortcuts. When deadlines are tight, professionals default to the easiest path — which is often "just ask AI for a full post." Documented workflows serve as forcing functions that maintain quality under pressure.

  • Speed gains are real, but the source of quality is still human. AI compresses the mechanical work of writing. The source material, the brief, the judgment about what is true and important, the voice, and the final quality review remain irreducibly human contributions.