Progressive Elaboration - Self-Publishing 4D PKM in 6 Weeks - VLOG Episode 5

A self-publishing sprint is turning into a workflow lesson: writing a full book with GPT works best when drafts are generated in small, connected units—then stitched together using “progressive elaboration” rather than trying to produce long, cohesive chapters in one pass. After roughly 20 hours of weekend writing, the author had reached 35,853 words in an AI-assisted first draft, but the text quality exposed a practical bottleneck: large prompts often produced good snippets that failed to align with each other, leaving gaps in chapter-to-chapter cohesion.

The process began with a reader-centered transformation map. The author defined who the reader is at the start and who they become by the end, then created multiple reader “types” with paired paragraphs for each. From that foundation came a detailed outline: an intro, multiple parts, an epilog, and chapters broken down into sections and bullet points—an outline that ballooned to about 10,000 words at the section level. Feeding that outline into GPT produced engaging prose, but not consistently cohesive long-form output.

The key constraint was scale. When GPT was asked to generate an entire chapter—roughly 3,000 to 4,000 words—the model struggled to maintain cohesion and the author found performance dropped sharply beyond around 8,000 words of generated material. Even when the text was later expanded in an editor, it rarely exceeded about 1,500 words reliably. The workaround—generating chapters section by section—solved the length problem but introduced a new one: each section read well on its own while the chapter’s internal continuity and overall narrative “fit” suffered.

To fix that, the author redesigned the prompt structure around transitions. For each major book segment (intro parts 1–4 and the epilog), GPT was tasked with producing three anchor statements: an opening paragraph, a closing paragraph, and a transformation description tied to that segment’s objective. Those anchors were then “exploded” downward: GPT used the segment-level transformation and outline to generate chapter-level opening/closing paragraphs and chapter content descriptions, and then repeated the same pattern at the section level. By giving GPT recurring context—stored in the opening and closing frames—it produced text that matched more naturally across boundaries, even if “rough edges” still remained and would require manual cleanup after a full draft read-through.

The author’s immediate plan is to finish the remaining parts quickly but not recklessly: complete part four and the epilog within the next few days, then take a holiday reading pass to mark inconsistencies and rough spots. Notes will be converted back into the draft with GPT, followed by a second read to tighten the final version. The goal is self-publishing by March 15, paired with a proper launch rather than a rushed release that could undermine sales. Alongside the writing workflow, the author also shared cover art experiments, weighing a more vibrant, engaging design against a cleaner, more focused alternative—seeking feedback on which better matches the book’s heart.