An audit pass across the catalog turned up the same defect in piece after piece: a core concept explained two or three times, multiple conclusion sections, a "Ghost Says…" coda restating the intro almost word for word. The writing itself was sharp — good examples, real technical depth — but the structure kept looping, the model saying one thing three different ways and calling the third one a summary.
It was all written by AI, and the repetition was consistent enough to be a signature rather than an accident. Which means it's fixable, but only if you understand where it comes from.
Training Data, Attention Heads, and Architectural Blindness
Language models train on internet text, and internet text is repetitive by design. Academic papers run abstract to introduction to body to a conclusion that restates the abstract. Blog posts run hook to explanation to takeaway to a summary that restates the hook. Business documents open with an executive summary and close by repeating it. The lesson the model absorbs from all of it is simple: say the important thing three times, at the open, in the body, and again at the end.
The attention mechanism compounds the habit. A transformer generates each token by looking at the tokens before it, tracking which words fit the context based on training patterns — but nothing in that process asks whether this point already got made three paragraphs ago. The model produces locally coherent text with no global awareness of its own redundancy.
Temperature pulls the wrong way too. Low settings favor high-probability tokens and safe, familiar patterns, which means more repetition; high settings buy variation at the cost of hallucination. Creativity, accuracy, and repetition-avoidance can't all be maximized at once. And the context window closes the trap: a long article overruns it, so the section written at token 15,000 has no memory of the near-identical section at token 3,000. The repetition isn't carelessness. It's blindness, built into the architecture.
The Five Repetition Patterns
Multiple conclusion sections. Introduction, body, then "What This Means," "The Bottom Line," "Ghost Says…," and a final wrap — each one restating the thesis under a new header. Read every section header after reading its body and ask whether the section added information. If it only re-emphasized, delete it.
Premise restated in the ending. The opening explains the core concept, the body gives evidence, and the closing section explains the core concept again. Compare the opening paragraph to the closing one directly. If they carry the same idea, the ending needs new context or it needs to go.
Concept explained twice "for clarity." A thorough technical explanation early, then a second pass later that re-explains the same mechanism more slowly. Identify the core concepts, count the explanations per concept, keep the strongest, cut the rest.
Redundant lists. One section lists a set of items; a later section lists the same items reworded under a different heading. Pull every list, compare across them, merge anything that overlaps.
Progressive scope narrowing. The subtle one. The intro states a broad idea, the next section states it slightly narrower, the next narrower still, and the coda calls it the key insight. It reads like progression but it's the same message at four zoom levels. Summarize each section in one sentence; if the summaries are variations on one thesis, merge or cut.
Red Team Prompting Techniques
The fix is to prompt against the architecture instead of hoping the model self-regulates.
