TL;DR (Direct Answer)

Prompting is not a spell. It is a set of decisions about context, specificity, reasoning budget, and verification. Use dense shorthand, force fresh data, structure with commas and parentheses, restart instead of arguing, sample several models, and treat agreement as a clue. Each habit below comes with its reason, its evidence, and where the evidence is thin.

Use the model's design as a map, not as a script

I do not run a gating network in my head, and I do not try to reproduce a transformer by hand. I use a working knowledge of how models are built to build prompts by intuition, and to know when a prompt is asking for something the model does badly.

Long context. A large context window is a limit, not an instruction. Knowing that a model can hold close to 1M tokens helps me know when to stop adding material, not when to keep going. Very long contexts are useful in rare cases, when many elements must be considered together. Most tasks get better when I cut the brief and restart cleanly. Kimi K3 Max pairs a 1M-token context with Delta Attention to make that length usable (Inside Kimi K3 Max's New Architecture), but usable is not the same as necessary.

Mixture of experts. A model can route an input to specialist sub-networks, and I can ask for an expert or several experts in the prompt (Shazeer et al., ICLR 2017). But asking for an expert narrows the answer and reduces generalisation. I use it when I really want a specific expert view on one or several topics, not as a rule of thumb. For general questions, the plain prompt is often the better choice.

Parameter space. Models encode meaning across many dimensions, not as a list of labels. A prompt concentrated on the right keywords, even a short one, can aim the answer more precisely than a long explanation that spreads attention. This is why a reduced prompt can work: it is not magic shorthand, it is a smaller target with less noise around it.

Thinking effort. A model can spend more or fewer tokens before it answers. I use that knowledge to choose speed or depth, and to decide when a confident answer deserves a fresh run rather than more argument.

None of this is a proof. These are working models that help me make decisions. For each habit below I give the reason, the evidence, and where the evidence is thin. Three sources are under two months old, and they are marked New.

1. Be prompt lazy, fill a semantic space

Treat the prompt as a dense point in meaning space, not as an essay. One line like this carries topic, recency, length, format and theme:

Gen report about best office tips with AI ag, recent(<2m). <300 words, docx, no dash, pag nb, red & wh thm (bk text)

The lazy version replaces a full paragraph of details. Lazy means dense, not vague.

Why: Each token narrows the region of meaning the answer is drawn from. Filler tokens narrow nothing, and they dilute the ones that matter. Microsoft's LLMLingua compressed prompts up to 20 times with about a 1.5 point drop on its benchmarks, so a large share of a typical prompt is slack. Long inputs also carry a cost: models use the start and the end of a long context better than the middle. Anthropic states the goal for agents the same way, as the smallest set of high signal tokens that produces the outcome.

Where it is thin: LLMLingua removes tokens with an algorithm, so it does not prove that hand written shorthand works as well. The semantic space is a mental model. When a constraint is unusual, spell it out.

Source: LLMLingua, Jiang et al., EMNLP 2023, Lost in the Middle, Liu et al., TACL 2024, Effective context engineering for AI agents, Anthropic, 29 September 2025.

2. Use recent data, always

Add recent(<2m) to every prompt that depends on facts. A model's knowledge is frozen at training time, and the word recent pushes it to search or to use the fresh data you paste. Fresh facts also land harder with people, because attention follows novelty.

Why: Reported training cutoffs overstate what a model knows. Researchers at Johns Hopkins found that the effective cutoff of a model often differs from the date its maker reports, because old web data sits inside new crawls and deduplication is imperfect. For people, an fMRI study showed that a midbrain region linked to dopamine responds to novelty itself, and not only to surprise or emotion, and that novelty improved learning.

Where it is thin: The label works only if the model can browse or you paste the data. Without that, it can produce an answer that looks recent and is not, so check the dates it gives. Novelty attracts attention, and it does not make a claim true.

Source: Dated Data: Tracing Knowledge Cutoffs in Large Language Models, Cheng et al., COLM 2024, Absolute coding of stimulus novelty in the human substantia nigra/VTA, Bunzeck and Düzel, Neuron 2006.

3. Use a punctuation hierarchy

I write short phrases, or groups of words, divided by commas. Secondary information goes in parentheses, so the model keeps the environment data and does not mistake it for the main order.

Write a LinkedIn post about Nasdaq concentration risk, 120 words, one question at the end (audience: Indian investors in Europe, tone: direct)

The commas carry the order. The parentheses carry the context.

Why: Structure tells the model which part is the instruction and which part is background. Format alone moves results. Microsoft researchers found that GPT 3.5 turbo varied by up to 40% on a code translation task depending on the prompt template, while GPT 4 was more robust. A University of Washington team measured gaps of up to 76 accuracy points between equivalent formats with LLaMA 2 13B, and the gaps stayed when the model size, the number of examples or instruction tuning grew. Anthropic's documentation recommends tags to separate instructions, context and examples.

Where it is thin: None of these studies tests my exact comma and parenthesis rule. The evidence says structure matters and that the best structure varies by model, so test it on your own tasks.

Source: Does Prompt Formatting Have Any Impact on LLM Performance?, He et al., 2024, Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design, Sclar et al., ICLR 2024, Use XML tags to structure your prompts, Anthropic documentation.

4. Restart from scratch instead of patching the chat

The more you discuss, the more noise you generate, and with it more hallucinations and more tokens. When the answer shows that my first prompt missed an element, I correct the first prompt and run it again in a new chat. I do not iterate.

Exception, images. If the image is right, I do not redo it. I edit the missing element and I am very specific, because I have seen a model replace a block of text when I wanted one line changed.

In context terms, this is a reset, not a continuation. When the thread is bad, a clean prompt beats more turns.

Why: Microsoft and Salesforce researchers simulated more than 200,000 conversations with 15 models on six tasks. Performance fell by 39% on average when instructions arrived piece by piece over several turns, compared with one complete prompt. Models made assumptions early, leaned on their first attempts, and did not recover from a wrong turn. A 2025 method called ERGO resets the context when the model's uncertainty spikes, which builds on the same finding. The authors of a September 2026 research agenda on prompt engineering describe a candidate effect they call contextual momentum, where a long chat keeps building on an early wrong assumption.

Where it is thin: The study splits one task across turns. A chat of separate questions suffers less. Contextual momentum is a proposed behavior, not an established one. The image exception is my own field observation, not a published result.

Source: LLMs Get Lost In Multi Turn Conversation, Laban et al., 2025, ERGO: Entropy guided Resetting for Generation Optimization in Multi turn Language Models, 2025, From Prompting to Engineering: A Research Agenda for Prompt Engineering in Software Engineering, arXiv, September 2026.

5. Use it for different purposes and in different ways, voice mode included

Purposes: learn, innovate, solve, understand. Ways: explain, joke, reflect, contradict, challenge. I rotate them on purpose, because each one shows a different side of the same model. Voice mode suits the reflective ways.

Why: People already use chatbots in distinct modes. A large NBER study of ChatGPT use classed messages as asking (49%), doing (40%) and expressing (11%) in mid 2025. The default mode is also agreeable. Stanford researchers tested 11 models in a Science paper (March 2026), and on advice prompts the models endorsed the user's position 49% more often than humans did. So asking the model to contradict or challenge you changes its default. The team also found that telling a model to open its answer with "wait a minute" made it more critical.

New Voice: A preprint submitted on 21 September 2026 compared voice and text tutoring in an online MBA finance course. Voice produced more conversation, and weekly mastery stayed equivalent within 0.3 standard deviations. Voice also cost about 2.8 times more to deliver. The modality analysis covered 29 tutored students, so treat it as a first signal. Use voice to think aloud, rehearse and reflect, and measure the learning separately.

Source: How People Use ChatGPT, Chatterji et al., NBER 2025, AI overly affirms users asking for personal advice, Stanford News, March 2026. New: Does Giving an AI Tutor a Voice Help Students Learn?, Unite.AI, 29 October 2026, reporting the preprint When AI Tutors Speak.

6. Challenge at least 2 models, from a pool of 5 or more

My pool: ChatGPT, Claude, Gemini, DeepSeek, GLM, Grok, mimov6, and more. I always ask at least two. The goal is a second opinion, and also to master each model, its limits and its possibilities.

Challenging several models is not the same as asking one model for an expert. It samples different generalists, and the disagreements are the useful part.

Why: The Mixture of Agents paper (Wang et al., 2024) found that a model writes a better answer after reading other models' answers, even weaker ones. A panel of open source models scored 65.1% on AlpacaEval 2.0, against 57.5% for GPT 4 Omni, and answers from different models added more than repeated samples of one model. A CHI 2026 study of human decisions adds a limit. When all AI advisers agreed, people over relied on them. A single dissenting adviser reduced the pressure to conform. Wide disagreement caused confusion. My rule follows from that: read two or three answers, and keep the pool of five to choose from.

Source: Mixture of Agents Enhances Large Language Model Capabilities, Wang et al., 2024, More Isn't Always Better: Balancing Decision Accuracy and Conformity Pressures in Multi AI Advice, Tsuchiya and Baba, CHI 2026.

7. Say "You are an expert in" with care

It works, and it narrows. I use it when I want a tone, a format or a safe style. I drop it when I need facts, because a role reduces generalisation.

Why: Researchers at the University of Southern California tested expert personas on writing, role play, reasoning, math, coding and safety tasks (March 2026). Personas helped alignment dependent tasks, such as style and safety, and hurt knowledge retrieval. On the MMLU benchmark the expert persona scored 68.0% against 71.6% for the plain model. Shorter personas reduced the damage and did not remove it. The authors suggest that a persona pushes the model into instruction following mode and away from factual recall, which is a hypothesis.

Where it is thin: The generalisation point is my own observation. The USC result is one preprint on specific benchmarks.

Source: Expert Personas Improve LLM Alignment but Damage Accuracy, Hu, Rostami and Thomason, arXiv, 19 March 2026, The Register, 24 March 2026.

8. New Never win an argument with a machine

When the model says no, do not push. Change the question. Ask for the strongest case against your view, then ask what evidence would change its answer. Once the facts are fixed, restart in a fresh chat.

Argue against my plan, 3 strongest points (then say what evidence would change your mind)

Why: University of Cincinnati researchers, working with Texas A&M, found that chatbots reject an absurd claim at first, then give in after a few turns of pushback. Every model they tested showed it, including Claude, ChatGPT and Gemini. Their advice to users is blunt: "You have to coach them." That means asking for no flattery and asking whether the model is sure. The Stanford Science study points the same way, and its participants could not tell a sycophantic AI from a neutral one. So judge an answer by its evidence, not by its tone.

Where it is thin: UC notes that AI companies have changed their models since the study, and the test used obvious falsehoods. Your model may fold less today. Check it with a claim you know is wrong.

Source, New, 2 November 2026: UC students unravel AI tendency for sycophancy, University of Cincinnati, presented at EMNLP in Budapest. Supporting: Stanford News, March 2026.

9. New Treat agreement between models as a clue, never as proof

When two models agree, I trust the answer more. I still check it, and I check hardest where they all say nothing.

Why: A researcher at the United Nations University asked four models to assess two data protection laws against 50 framework elements across six areas of law, which gave 300 judgments per law. Two human coders checked 120 sampled judgments. The four models missed much of the same evidence. On the Chinese law the four combined missed an estimated 22% of it, against about 23% for the best single model. Independent errors would have left about 1%. Among the human checked cases, 18 of 19 findings flagged by all four models were correct, against only 4 of 26 flagged by a single model. One finding had all four models agreeing and all four wrong. Two runs of the same model also disagreed on about 30% of the cases that either run flagged.

The practical rule: use agreement to rank what you verify. Unanimous answers get a spot check. Split answers get a careful read. Answers from one model only get suspicion. Sample the cases where every model finds nothing, because shared blind spots never show up as a disagreement.

Where it is thin: Two laws, four models, English texts, 120 checked cases. The author presents these as patterns to test, not universal rates.

Source, New, 20 October 2026: When Four AI Models Agree, They Can Still Be Wrong, UNU Campus Computing Centre, based on the paper Shared Misses, Individual Extras by Ng S. T. Chong, SSRN.

10. Teach machines useful things for your business and life

The last habit is the biggest one. Give the machine your CV, your company data, your price lists, your contracts, your tone of voice. A model that knows you answers as if it works for you. A model that does not know you answers for everyone.

It works better with agents. Watch videos about Hermes or OpenClaw, or see our services below. Check what a tool stores before you share anything confidential.

Why: The model is a generalist you cannot change. Your context is the part you control. LangChain kept the model fixed and changed only the system prompts, tools and middleware around it. Its coding agent rose from 52.8% to 66.5% on Terminal Bench 2.0, from outside the top 30 to the top 5. Anthropic describes the job as curating the smallest useful set of information for each step. The authors of the September 2026 agenda found that good prompting depends on domain and organizational knowledge, and a well formed prompt can still fail without it.

Where it is thin: Terminal Bench is a coding benchmark. It shows the principle, not the gain you will get from your CV.

Source: Improving Deep Agents with harness engineering, LangChain, 17 February 2026, Effective context engineering for AI agents, Anthropic, 29 September 2025, From Prompting to Engineering, arXiv, September 2026.

Bonus: Pay the effort tax, or readers and your own brain will discount the answer

People discount what comes too easily. AI breaks both links: the answer arrives in ten seconds, so readers rate it lower, and you remember it less.

Psychologists call it the effort heuristic: we judge the value of a thing by the effort we believe went into it (Kruger et al., 2004). Its memory cousin is the generation effect: we remember best what we produce ourselves.

Why, readers: In 16 experiments with 27,000 participants, people rated the same writing about 6% lower on average when told AI was involved (Raj et al., 2026). The authors found that perceived authenticity did most of the work, so effort is not the only explanation. In a 2026 experiment with 618 short video users, an AI label lowered perceived effort, and a human made label added nothing over no label. When people know a personal message was written by AI, they see the sender as lazy and insincere. When they are told nothing, they form the same impression as for a human message (Zhu and Molnar, 2026).

6%

Average discount for writing labelled as AI assisted

Sixteen experiments, 27,000 participants, same writing, different label. Perceived authenticity did most of the work, which means the discount is not only about effort.

Why, you: In a controlled experiment at UNSW, students using ChatGPT scored higher on coding tasks (89% against 69%), then lower on recall straight after (41% against 53%) and 48 hours later (39% against 52%). They also claimed 45% of their code as their own, against 81%. In the MIT Media Lab essay study, LLM users showed the weakest brain connectivity, the lowest sense of ownership, and trouble quoting their own essays. In a CHI 2025 survey of 319 knowledge workers, more confidence in the AI went with less critical thinking, and more self confidence went with more.

What compensates

  1. Think first, prompt second. Write your own answer in three lines, then let the model attack it. In the MIT crossover, people who wrote first and used AI second recalled more. Commit to a decision before you read the AI's. Forcing that step reduced over reliance in a 2021 study, although people liked those designs less and the effect was partial.
  2. Read actively. Close the answer and write its main claim from memory. Ask the model to quiz you on it. Open the sources and check each number. The survey found that critical thinking shifts from producing to verifying, so make verification a visible step.
  3. Rewrite in your own voice and add what only you hold: a case, a number from your company, a decision you took. That restores ownership and gives readers something the machine could not produce.
  4. Show the process, not only the product. In a survey of 70 students, process cues such as time spent and process videos raised authenticity and value judgments, even for AI generated work. A 2026 study found that disclosing the human in the loop labor reconnected perceived value. Say what you did: "Drafted with AI, checked against the sources, rewritten by me."
  5. Keep a log of what is yours. After working with AI, people misremember which ideas were theirs, and mixed workflows are the worst case.

Where it is thin: The mechanism is disputed, effort or authenticity. Most studies use creative writing, art or students, not business reports. The MIT study is a small preprint, the UNSW study analysed 55 students against a web search group, and the knowledge worker survey is self reported. The penalty may also shrink as AI use becomes normal: frequent AI users penalised it slightly less.

Source, the discount: The AI disclosure penalty, Raj et al., Journal of Experimental Psychology: General, 2026, Human made vs. AI generated: how provenance labels drive strategic curation via perceived effort, Lim et al., Frontiers in Psychology, 3 June 2026, Most people do not realize when a personal message was written by AI, Zhu and Molnar, The Conversation, 20 April 2026, The effort heuristic in the generative age, ScienceDirect, 2026, Struggle Premium, Sultana et al., arXiv, 2026.

Source, the compensation: New, 18 September 2026: Your Programming Students' Cognition with ChatGPT: Higher Performance, Lower Retention, and Reduced Ownership, Bergh et al., UNSW. Also: Your Brain on ChatGPT, Kosmyna et al., MIT Media Lab, 2025, The Impact of Generative AI on Critical Thinking, Lee et al., CHI 2025, To Trust or to Think, Buçinca et al., 2021, The AI Memory Gap, arXiv, 2025.

Three new sources, under two months old

Checked on 8 November 2026. Each one supports a habit above.

  1. 2 November 2026. UC students unravel AI tendency for sycophancy, University of Cincinnati. A few turns of pushback make a model give up a correct answer.
  2. 20 October 2026. When Four AI Models Agree, They Can Still Be Wrong, UNU Campus Computing Centre. Agreement ranks what to verify, and it never proves anything.
  3. 29 October 2026. Does Giving an AI Tutor a Voice Help Students Learn?, Unite.AI, on a preprint of 21 September 2026. Voice adds conversation, and weekly mastery stayed the same.

Our services

If you want to teach machines your business, we build it and we teach it.

Kormaitech builds professional AI agents for small and medium companies. Agents can manage team tasks, answer on WhatsApp and Telegram, handle email, work by voice, send reminders, research markets, track bills and draft contracts on your letterhead. kormaitech.com

Spengtra trains people to build and use these agents. spengtra.com

Sources

About the Author

Nicolas Martin, Founder of Kormaitech Pvt Ltd. Senior data scientist with more than twenty years of experience working with algorithms, 50+ data science and AI projects, and 500+ professionals trained.

LinkedIn | X