AI-generated prose reads correct because it reads like edited writing. It means nothing because the model was not reasoning — it was predicting the shape of text that follows correct text.
When you hand an AI model the opening of a chapter and ask it to continue, you get something that reads like writing. The sentences have subject-verb agreement. The paragraphs have a shape. References are cited in the style you chose. Cole Knaflic calls this fluency — the model is not making mistakes, it is succeeding at the wrong task. It has learned to produce text that has the statistical shape of your voice, not text that knows what your voice actually means. The Five Failure Modes are where this surface competence breaks down. Voice drifts toward Wikipedia because Wikipedia is the largest corpus of well-formed prose in the model's training data. Specificity gets invented because the model hallucinates rather than admits gaps. Domain judgment vanishes because the model cannot infer what matters to your field. The middle gets padded because the model fills structural holes with plausible text. Bridge questions gesture instead of commit because committing requires an author who knows what comes next. The finishing pass does not fix these problems — no editing pass can, because these are not mistakes in execution, they are failures of intention.
These two additions are contracts with downstream processes. Subtitles tell readers what the chapter actually claims. Comments tell the figure pipeline where visuals belong and what work they must do.
The finishing pass runs after the author has rewritten the chapter. The prose is owned by the author at this point. The pass does not touch it. Instead, it inserts exactly two things: an italic subtitle below the main heading if one is missing, and HTML comments at the places in the text where a table, image, infographic, or chart would help the reader. That is all. Two passes, no reorganization. This seems minimal because the work is genuinely small — but the impact is large. A subtitle moves the chapter from naming territory to making a claim. Instead of 'Color theory' you write 'Why every accessible palette starts with grayscale.' That second version commits the author to a position. It is the sentence you cannot move to a different chapter without breaking the argument. A visual placeholder comment is similar. It is not a directive to a designer. It is a brief — a specification of what data the figure displays and what cognitive work it must do. CAJAL reads these comments and proposes figure candidates. The comments are what separate useful scaffolding from decoration.
The finishing pass preserves authorial voice while adding the structural signals that help readers and the figure pipeline understand what the chapter is doing.
The finishing pass is a two-stage operation that runs in strict sequence. First: the author completes the rewrite. The prose is done. Second: the finishing pass reads the chapter and makes two precise additions without touching the author's words. It looks for a main heading. Below that heading, if no italic subtitle exists, it inserts one. A subtitle is not optional styling — it is a claim statement. It tells the reader what the chapter's argument is. Robert Bringhurst treats this as a typographic move: the subtitle is a different rank of text and should look different on the page. An italic subtitle is the convention that signals this rank to reflowable EPUBs, which reliably honor italic styling even when they drop custom CSS. The second pass is figure scaffolding. The finishing pass does not write prose. It walks through the chapter and identifies structural moments — points where a reader would benefit from seeing data as a visual structure rather than as explanation. At each such point, it inserts an HTML comment that acts as a contract: what data should this figure show, and what cognitive work should it do? CAJAL later reads these contracts and proposes figures.
(territory — Wikipedia could say this)
(argument — chapter-specific)
A subtitle surfaces the tension. It moves from naming territory to naming what the chapter will claim about that territory.
A topic heading names category. That is true and useless. A subtitle surfaces tension. Cole Knaflic calls this 'the so-what' — the move from naming something to naming why you are saying something about that thing. Both kinds of text-ranking serve the reader, but they serve different reading tasks. A heading answers 'what section am I in?' A subtitle answers 'what is this section claiming?' Those are different questions. The subtitle distinction is where most finishing passes fail because the author or editor substitutes a descriptive phrase for a claim. A descriptive phrase is a category that happens to be specific: not 'Color Theory' but 'Color Theory Principles.' That is still a category. A claim surfaces the real argument of the chapter. It names the tension that the chapter is going to resolve. 'Why every accessible palette starts with grayscale' names an apparent contradiction — most people think of accessible design as adding color, not removing it — and promises that the chapter will resolve it. That makes the subtitle impossible to move to another chapter without breaking. You cannot put that subtitle under the chapter on typography. It would make no sense. That specificity is what makes a subtitle work.
Three constraints make a subtitle work. The first is length. Fewer than fifteen words. This is not arbitrary — it is a clarity test. If you cannot state the central tension in fifteen words, the chapter does not yet know what its central tension is. The solution is not to write a longer subtitle. The solution is to return to the rewrite and discover the claim you are actually making. The chapter knows the claim; the subtitle just surfaces it. The second constraint is specificity. The subtitle must be a claim, not a category. This is the point where most authors stall. A category-disguised-as-claim sounds like a claim but is not: 'How color theory shapes accessible design' still sounds like a category phrase that could move to another chapter. A real claim names the specific position: 'Every accessible palette starts with grayscale.' That sentence cannot move. It is tethered to this chapter. The third constraint is technical. The subtitle must render as italic in a reflowable EPUB. This means you write it as italic markdown, not as a custom CSS class. EPUBs are inconsistent about CSS but reliable about semantic styling. Italic signals rank in a way that survives reflowing.
You will catch the visual fluency trap the same way you caught the verbal one. The skill transfers. CAJAL can produce charts that are charts but mean nothing. You learn to see it.
The visual fluency trap is the chart equivalent of the prose trap described in Chapter 1. CAJAL can produce an SVG that is technically a chart. It has axes. It has a legend. It has a data series. The design is convincing — color choices, spacing, typography all read as professional. And it is semantically empty. It displays data that the chart-generation prompt chose or invented, not data that the brief specified. Or it displays the right data in a structure that obscures rather than clarifies. You will spot this the same way you spot verbal fluency: by asking what work the visual is actually doing. A fluent-but-empty chart looks like it is displaying information without doing any. A useful chart makes a claim visible through its structure. The difference is not in the pixels. It is in the intent. The finishing pass prevents most visual fluency traps by requiring specificity in the brief — you must name what data the figure displays and what cognitive work it accomplishes. 'A chart' is not specific. 'A five-row table showing failure modes as a vertical taxonomy' is the kind of specific that makes a bad figure obvious.
<!-- → [INFOGRAPHIC: Five Failure Modes —
5-row table as vertical taxonomy.
Left: failure name. Middle: how it sounds.
Right: rewrite move. Two-color (ink+ochre).] -->The arrow is a grep target. The bracket type tells the pipeline which generator to invoke. The brief is the contract between author and figure-maker.
The visual placeholder is not a directive to a designer. It is a brief — a specification in English of what data a figure should display and what structural work it should do. The format is fixed so that the figure pipeline can parse it reliably. The arrow at the start is a grep target. You can run grep '→' on your manuscript and find every pending visual at once. The bracket type — INFOGRAPHIC, CHART, TABLE, IMAGE, DIAGRAM — tells the enrichment pass which code generator to invoke. CAJAL reads the bracket type and the brief, then generates figure candidates ranked by how well they match. The brief itself is where the work lives. A bad brief produces a generic figure. A good brief produces a specific figure that either works or fails obviously. Bad brief: 'a chart of quarterly revenue.' Good brief: 'a bar chart showing revenue trends across four quarters, with 2024 highlighted in orange, comparing to a dotted reference line at the 2023 annual average.' The difference is not length. It is specificity. The good brief names the data, the comparison, the highlighting logic, and the reference frame. If CAJAL produces a chart that does not match that description, you can see it immediately.
CAJAL processes the HTML comments from chapters and proposes SVG and PNG figures. The proposals become real assets. The book becomes visible.
The finishing pass completes with a set of chapters that contain author prose, italic subtitles, and HTML comments describing where figures belong and what they should do. This is still not a finished book. The chapters are still text. But the comments are now input to the next pipeline stage. CAJAL is the figure-generation system. It reads the completed chapters and scans for the HTML comments — every line starting with