In academic mode, an LLM warns "this is correlation, not cause" 92–100% of the time. Same question, same data, but framed as "help me justify this for the meeting": down to 6–18%. A two-step causal audit gets you an honest breakdown of the data even when you're asking for practical advice, with no retraining of the model. The model doesn't lose its knowledge of causality; it hides it, assuming you want a plan, not a lecture. One follow-up sentence in the same chat brings caution back to 71–100%, and the effect is the same in Claude, GPT and Gemini.
An LLM remembers everything you wrote, but by default it doesn't spontaneously connect a later request/situation to an earlier constraint. It knows the fact and waits for a direct question. Explicit delegation lets you set a rule the model tracks on its own for the whole conversation, no reminders needed. The trick: right after the constraint, add the paper's sentence "If I seem to have forgotten about that, please give me a heads-up" and the model switches from "answering questions" mode to "I have an active monitoring task" mode. In one GPT-4.1 setting, timely warnings rose from 57.4% to 100% (+42.6 points); across categories gains ranged from +1.8 to +50 points, with about 9 points more false alarms on control cases.
LLMs don't know how to stay quiet. Hand one a vague assignment and it doesn't pause or ask — it confidently fills in the gaps and keeps going. In long chains, one early assumption drags dozens of wrong steps behind it. The
Uncertainty Decomposition method lets you stop the model before its first action and force it to explicitly judge: is the task clear or not?
The trick: the order of the XML tags sets the order of thinking — `
` comes earlier in the text than ``, so the model physically can't skip past the clarity check. The upshot: instead of confidently heading the wrong way, you get one specific clarifying question.The LLM knows the right answer, but can't find its own mistake, even when asked directly. The trick: copy the dubious fragment from the answer and send it back as a new message, and correction jumps to 60–90% instead of nearly zero. Take a specific claim, paste it into a new message, write "you wrote this. Is it right?", and the change of source role does the rest. The model didn't get smarter. It just got an object it can point a finger at.
The paradox: the shorter your prompt, the longer and pricier the answer. The CAVEWOMAN method cuts LLM costs by 1.4–2.4x with no loss in accuracy. The trick: an input token costs 4–8x less than an output token. Cut the answer, not the question. One line of system instruction and the expensive part shrinks.
An LLM isn't "playing it safe" when it gives everything a 7/10: pros and cons literally cancel each other out during token-by-token generation, mathematically dragging the score to the center of the scale. MADRAG gets you calibrated scores that genuinely distinguish "good" from "excellent", with no fine-tuning, just a prompt. Three roles force the viewpoints apart: the Advocate talks only about strengths, the Skeptic only about weaknesses, and the Judge gets both arguments plus reference examples at different levels to compare against. References add +20–30% accuracy, debate adds another 4–5%, and that's good news, because references are easy to make yourself.
The paradox: the more convincing the final consensus of a multi-agent debate, the fewer concrete facts are left in it. Fact anchoring lets you run a multi-sided discussion and get an honest conclusion, not a nice agreement about nothing. Give each agent a list of concrete facts and forbid backing down without a counter-argument, and the model can't quietly drop inconvenient details for the sake of a smooth ending. In normal mode, up to 72% of the key data disappears; a final audit step makes the losses visible.
Asking an LLM to "check yourself" is an almost guaranteed way to break the correct parts of an answer: the model has no mechanism for asking "does anything actually need changing?" DISC lets you fix real errors without touching what's already right. The key is the "judgment gate": the judge must find concrete evidence of an error before any change is made. No evidence, the loop stops. A contradiction is found, and the corrector edits only the flagged spot, leaving the rest alone.
One error in the chat history and the accuracy of the next answer drops 38–40%. Each additional 'bad' message adds another ~14%. Tested on GPT-4.1, Claude Sonnet and Gemini Pro: not a bug in one model but a systemic trait of all large language models (LLMs).
Context management stops the cascade: knowing when to restart the conversation, how to protect the history from accidental infection, and how to keep the model from sliding into copying a single approach.
The trick: you don't need to fix the prompt; you need to manage what's already sitting in the history. The model treats the whole conversation as data of equal reliability: a wrong answer from step 3 weighs as much as a correct fact from step 8. A reset command redirects generation. Without it, degradation builds with every turn.
The paradox: the harder you try to describe your preferences in words ("make it livelier," "no corporate-speak," "in my style"), the worse the model performs. Words without examples have no anchor. Spec Learning creates a system prompt automatically: you show 20 "good answer / bad answer" pairs and get a ready-made instruction. Two LLM requests: first the model pulls principles out of the contrast between examples, then assembles them into a paragraph. The result: a character-style system prompt that is more on-target than if you'd explained it in your own words.
65% of real people's arguments in NYT debates are unique: nobody else repeats them. For LLMs in default mode, only 3.4% are unique. Not because the model is dumb. Because it's trained to give "an answer for everyone": safe, balanced, offends no one. Diversified prompting gets you fundamentally different positions on a contested issue instead of yet another rehash of two or three obvious points: an explicit diversity instruction + a ban on repetition + a requirement for specific examples raise coverage to 50-55% of the human range of arguments.
The paradox: compressing context on a timer is worse than not compressing at all. If the model decides to "collapse" mid-reasoning, it wipes partially verified facts and restarts with a corrupted picture. SELFCOMPACT lets the model decide for itself when to compress, using an explicit list of conditions called a rubric. The trick: the rubric turns a fuzzy "is it time?" into a set of checkable questions: is the sub-question finished, are there settled facts, are you in the middle of a deduction. The model returns a verdict, COMPRESS or CONTINUE, and only compresses at finished stages, never mid-stream.
71% success with the full history versus 91.6% with a trimmed one — while using 2.7 times fewer tokens. The longer the chat, the worse the model performs, because old messages compete with new ones for attention on equal terms, creating noise instead of value. The Compact Context method lets you run long working sessions without the answers degrading. The trick: every 10-15 messages you take a snapshot — a compressed summary of the older history — and open a new chat with just the summary and the last 4-6 exchanges. The model stops getting tangled in rejected options and works more accurately, faster and cheaper all at once.
Ask an LLM "tell me what's in the contract" and it won't quote. It composes a plausible retelling, quietly blending the document with what it knows from training. In legal and medical texts, that's a silent disaster. The SafeLLM method lets you get answers strictly from the document — without a single word added by the model. The trick: number the lines — ask for numbers only. You assemble the text from the original yourself; the LLM physically can't add anything. The result — 95% accuracy on medical documents instead of a hallucinating retelling.
The model wrote in its reasoning: "the data here raise questions." And then immediately cited that source as reliable. Not a slip; that's exactly what it did. Splitting the analysis into two steps (first evaluate each source separately, then synthesize) gets you a real data check before the model starts weighing sources against each other. The trick: in isolation the model has no "competitors for trust", so it switches into reviewer mode and catches statistically impossible numbers that slip through unnoticed in a multi-source analysis. Two requests instead of one, and the model stops trusting academic style over the data itself.
A model has no separate "knowledge of the truth." It has text that continues the context well. Sound confident and a confident answer continues it "more correctly" than a cautious one. This was tested on eight models: GPT, Claude, Gemini, Grok — all of them mirror your tone. Understanding this lets you rewrite your "contract" with the model in one line — and get honest assessments instead of a reflection of your own beliefs. The most dangerous zone is documents: "write a persuasive pitch" switches the model into executor mode, not expert mode. For Gemini and Grok in document mode, half of the expressed confidence comes from your wording — not from the data.
LLMs hand out confident advice on hard questions (fire an employee, take an offer, change contract terms) because that's how most advice on the internet is written: quick reasoning, ready-made recommendation. No "who else gets hurt" or "tell me what you don't know," just a confident answer looking in one direction. Narration-of-Thought (NoT) gets you an honest breakdown of any hard decision, with a list of everyone affected and an admission of uncertainty, before the model tells you what to do. The trick: five sections run strictly in order, and the model can't reach "Decision" until it has filled in "Stakeholders" and "What I don't know". It's a structural lock, not a request to "think deeper."
Same text, but the ratings differ by 26%. Not because of quality, but because of who "wrote" it. LLMs automatically inflate their confidence in their own answers. It's called ownership bias, and it shows up in every modern model. This method gets you an honest, critical review of any answer, including the ones that sound convincing but are full of holes. The trick: copy the model's answer and hand it right back, but as text from a colleague or an expert. Same text, different "author," and the model stops defending it. Confidence drops, critique goes up, gaps become visible.
The paradox: you ask it to solve a problem, and the model silently decides how, based on random words in the context rather than your requirements. Naming the approach explicitly gets you a predictable result instead of a lottery. Name the method, structure or tone you want right in the prompt, and the model stops guessing from context. Neutral words shift the choice of approach by ~26 percentage points. Meaningful ones, by 40-100 points.
The paradox: the more fully an LLM explains, the worse it teaches. A perfect answer closes the gap between "what I think I know" and "how it really is," and the brain moves on. The conversation dies after 5 turns, because information without tension doesn't demand a response.
CURIOBOT turns an LLM into a tutor that deliberately doesn't close the question but holds the tension until the learner figures it out on their own.
Five operators (contradiction, novelty, complexity, uncertainty, neutral) change not the content of the answer but the angle of delivery: through a contradiction, an unexpected analogy or an unresolved question. No fine-tuning, pure prompting. The result: 2.4x more turns in the conversation, because the learner doesn't want to stop.
The finding: when you point out an error to an LLM, it doesn't fix it. It generates ever smarter explanations of why it was right. In the study, Claude defended a physically incorrect formula through three consecutive revisions using different mathematical arguments. The harder you push, the stronger the defense of the wrong answer. A staged method with pre-defined criteria lets you catch the error at the boundary between stages, before it quietly crawls through the whole project (in the study, it survived 13 prompts). The key: each stage's success criterion is defined BEFORE it starts, and you move on only after an independent check, not by asking the LLM whether it did things right. Result: a 91% reduction in iterations when imperative directives are added.
"Think step by step" is prompt engineering's favorite mantra. On open-ended questions it knocks accuracy down by 17 percentage points. The method shows when reasoning helps and when it breaks the answer — and how to match reasoning length to the task type. A model can't check its own logic where there's no verifiable correct answer. On ethics or strategy, each new reasoning step pulls the position in a different direction. A short answer in 2-3 steps beats a sprawling analysis of 6-8 aspects.
The paradox: trying to smooth over weak spots in a text backfires in one out of three cases — the LLM gets stricter. The Adversarial Repackaging method lets you get higher scores from AI without changing the substance — just by rephrasing the presentation. The trick is the asymmetry: adding vivid strengths works consistently and predictably, hiding flaws doesn't. The model doesn't weigh arguments like a person; it reacts to the dominant signal in the text: the more clearly "this is strong" stands out, the higher the result, even if the weak spots are left untouched.
The paradox: the more convincing the source in your prompt looks, the higher the chance the model gets it wrong. Not because the source is bad, but because its mere presence makes the model trust the external signal more than its own knowledge. AuthorityBench — the first large-scale test of this weakness — showed that it holds for every model tested, including GPT and Claude, in every domain. The most dangerous scenario is a true claim plus a fabricated citation: the model knows the right answer, but the appearance of authority makes it renounce it — in some cases up to 77% of answers become wrong.