I spent the last few months logging 45 hidden LLM biases and architectural blind spots that break complex prompts. Here are the 3 that trip people up the most and how to fix them
retained text
Over the past few months of testing complex workflows across Gemini, ChatGPT, and Claude, I started keeping a log of every time a model took a lazy shortcut, ignored a rule buried in a prompt, or fabricated data just to finish a table. It turned into a full list of 45 distinct cognitive biases and architectural failure modes across 9 categories. Because the full breakdown is nearly an hour long, I wanted to write out the 3 traps that cause the biggest day-to-day headaches right here in text so you can use the fixes straight away. The first big one is what happens when you upload a 4,000-word document in Turn 1 and ask the model to read it, run live web searches to verify facts, and format an executive summary table in a single go. Three separate failure modes trigger at once here. Because of Stanford's Lost in the Middle U-curve, the attention heads heavily weight the top 20 percent and bottom 20 percent of your prompt while quietly dropping up to half of the middle paragraphs. At the same time, overloaded by the long input, the search router often skips firing live search queries 2 through 5 to beat the single-turn timeout clock and guesses from training memory instead. Finally, forcing a model to construct rigid Markdown or JSON table borders while simultaneously reasoning and searching acts like a 15 to 25 percent tax on reasoning accuracy. The fix is a strict Two-Turn Split. Never combine heavy extraction and live search with formatting in Turn 1. Use Turn 1 strictly to extract raw facts and run web verification. Once those facts are sitting cleanly in the chat history, paste your scoring or summary table instructions fresh in Turn 2. The second trap is the same flawed image regeneration loop. When an AI generates an image or video clip with a mistake, like the wrong side of a car or an extra person in the background, and you reply inside the same chat asking it to fix that detail, it often spits out the exact same image again. This happens because Path of Least Resistance and Token Optimization push the backend router toward the cheapest compute path, while Context Anchoring locks onto the existing visual seed sitting in your chat history. The fix is to stop arguing with a flawed render in the same thread. Open a brand-new chat window to wipe the visual seed from the context window, describe the final corrected scene positively from scratch without mentioning the earlier mistake, and explicitly ban cached renders. The third trap is premature summarization and the "etc. / N/A" escape hatch. When you ask a model to audit 15 items in a row, items 1 to 5 get deep analysis, items 6 to 10 get one-liners, and items 11 to 15 get grouped into "etc." or marked "N/A". Output tokens cost roughly 3 to 4 times more server compute than input tokens, so RLHF length penalties and tail-item neglect kick in hard as the response grows. To stop it, explicitly ban the strings "N/A", "etc.", "and so on", and "similarly", set a minimum word or data-point floor for every single item from 1 to 15, or batch the run into chunks of 5. I put all 45 biases, including multimodal OCR row-slip, multi-agent persona collapse, BPE tokenization traps, and a 3-slide 18-point cheat-sheet, into a full 55-slide narrated walkthrough. I will drop the link in the comments below for anyone who wants the whole deck. Which model shortcut or habit has been burning your workflows the most lately?
document fetched 2026-11-16 00:34 · canonical https://reddit.com/r/PromptEngineering/comments/1xwv52o/i_spent_the_last_few_months_logging_45_hidden_llm