How Prompting Has Changed Since 2025
Imagine opening a prompt you saved in early 2025.
It likely starts with “You are a world-class expert.” Then comes a paragraph explaining the task, six rules, three warnings, a request to “think step by step,” some background copied from an earlier conversation, and a final reminder to “double-check everything.”
But in 2026, the AI models have evolved. Which means your prompting must evolve too. Current reasoning models are better at following instructions, planning complex work, using tools and working through problems internally.
The Psychology of Bad Prompts
Before rewriting your prompts, it helps to understand three ways a conversation can push an LLM toward worse answers:
1. The Attention Problem: More Context Is Not Always Better Context
Modern models can accept enormous amounts of information. That does not mean every piece of information receives equal attention.
Research into the “lost in the middle” effect found that models can perform worse when relevant information is buried inside a long context. In the original experiments, performance was often strongest when important information appeared near the beginning or end and weaker when it appeared in the middle.
Context is only valuable when it reduces ambiguity. Volume alone does not.
2. The Agreeableness Trap: AI Can Follow Your Opinion Into the Wrong Answer
Suppose you ask your AI: I think this campaign failed because our targeting was too broad. Analyze why broad targeting caused the poor performance.
You have already supplied the conclusion. The model now has two jobs competing with each other: objectively analyze the evidence and respond coherently to the premise you handed it.
This matters because sycophancy, the tendency of models to agree with users, is a documented problem. Anthropic research found that models can favor answers matching a user’s stated beliefs over truthful ones. More recent 2026 research has continued to find sycophantic behavior across current models.
Remove your conclusion from the question
Leading prompt | Neutral prompt |
|---|---|
Why is this strategy a bad idea? | Evaluate the strengths and weaknesses of this strategy. |
Explain why our pricing is too high. | Assess whether pricing contributed to the result. |
I think option A is clearly better. Do you agree? | Compare A and B against these criteria. |
Prove that this claim is wrong. | Check whether the available evidence supports this claim. |
When you need judgment, give AI the evidence and criteria before your opinion.
3. The Complacency Cost
Prompts tend to grow rather than shrink. AI ignores your format once, so you add: YOU MUST FOLLOW THE FORMAT EXACTLY.
It gives a shallow answer, so you add: Think deeply and comprehensively.
It misses an error, so you add: Review your answer three times before responding.
A year later, those instructions are probably still there. This accumulation is sometimes called prompt debt: instructions written to correct earlier model behavior remain embedded in prompts even after models or workflows change.
The result of this can be:
More input tokens
Longer responses
Unnecessary reasoning
Conflicting instructions
More tool calls
Harder-to-maintain prompt templates
The solution is not automatically making every prompt shorter. It is making every persistent instruction earn its place.
Prompting Habits That are Holding You Back
Here is the audit:
Prompting Habit | Why it was popular | What to do in 2026 |
|---|---|---|
Over-polite padding | AI felt conversational, so users wrote to it like a person | Be natural, but put the request first |
Context dumping | Bigger context windows encouraged “paste everything” | Curate context around the task |
Vague personas | “Act as an expert” was an easy way to steer responses | Define expertise, situation and responsibility |
Single-shot complex generation | One giant prompt felt efficient | Split complex work at meaningful checkpoints |
Leading the witness | Users ask AI to validate an existing belief | Ask AI to evaluate evidence before your opinion |
Zero output constraints | Users assume AI knows what “good” looks like | Define format and success criteria |
Advanced Prompting System for Current LLMs
You do not need another complicated prompting framework. For most everyday tasks, the upgrade is to replace old prompt habits with clearer instructions, better context, and constraints that tell the model what a useful result looks like.
Use this table as a reference when writing a new prompt or cleaning up an old one.
Instead of doing this | Do this instead | Why it works | Example |
|---|---|---|---|
Dumping instructions, context and examples into one block of text | Separate the prompt into clear sections such as Task, Context, Constraints and Output | Structure helps the model distinguish what it needs to do from the information it needs to use. recommends consistent prompt structure and delimiters such as headings or XML -style tags for complex prompts. | Task: Compare these three proposals. Context: [proposal data] Constraints: Budget is ₹5 lakh. Output: Comparison table + recommendation. |
Adding “think step by step” to complex prompts | Tell the model what needs to be checked or demonstrated | Current reasoning models can handle much of the reasoning internally. What you usually need is an answer you can verify, rather than a narration of every reasoning step. | “Recommend the strongest option. List the assumptions that affect your recommendation and cite the evidence supporting it.” |
Using vague personas such as “Act as a marketing expert” | Define the role, situation and responsibility | A specific role gives the model information that can actually change its judgment. | “You are advising a bootstrapped B2B SaaS company with one marketer and a $2,000 monthly acquisition budget. Recommend the most realistic channel for the next 90 days.” |
Giving AI every piece of background you have | Include context that changes the answer | Large context windows provide capacity, but research shows models do not necessarily use every part of a long context equally well. | Include customer research, budget and current positioning. Leave out an old meeting transcript unless it contains information relevant to the decision. |
Asking “What do you think?” after telling AI your opinion | Give the evidence and evaluation criteria before revealing your preference | This reduces the opportunity for sycophancy, where the model follows the user’s stated belief instead of independently evaluating the evidence. | “Compare A, B and C on cost, risk and expected return. Rank them before I tell you which one I prefer.” |
Asking AI to “be thorough” or “think deeply” | Specify where you need depth | Broad instructions can encourage unnecessary analysis. Specific requirements tell the model where additional effort is useful. | Instead of “Research this thoroughly,” use: “Verify the current price, cancellation policy and data- retention policy using primary sources.” |
Leaving the output open- ended | Define the format and success criteria | The model knows what task it is completing and what conditions the final answer must satisfy. | “Give me three recommendations in a table. Compare cost, setup time and privacy. Finish with one recommendation for a solo business owner.” |
Trying to complete a complicated project in one giant prompt | Create checkpoints where the output of one stage affects the next | You can catch weak assumptions before they become the foundation of the final output. | 1: Analyze competitors. 2: Review the 3: Identify positioning gaps. 4: Choose one. 5: Develop the positioning. |
Repeating an instruction because it seems important | State the requirement once and make it measurable | Repetition adds tokens and can make prompts harder to maintain without necessarily making the requirement clearer. | Instead of mentioning the word limit three times: “Maximum length: 800 words.” |
Using blanket “always” and “never” instructions | Explain when the rule applies | Conditional i nstructions preserve the rule without forcing it into situations where it makes the output worse. | Instead of “Never use bullet points”: “Use paragraphs for explanations and bullets for steps or lists that need to be scanned quickly.” |
Telling AI to verify everything repeatedly | Define what requires verification | Verification becomes tied to actual risk instead of creating unnecessary checking and extra work. | “Verify all prices, dates and current product features against primary sources before answering.” |
Keeping the same prompt after changing models | Retest saved prompts when you move to a substantially newer model | Instructions added to compensate for an older model’s behavior may no longer help and can sometimes make a newer model overthink or over-trigger tools. | Test the original prompt against a version with old reasoning instructions, repeated rules and generic warnings removed. |
A Simple Formula To Remember
Task + Context + Constraints + Output
If you want more practical guides for using AI at work without turning your job into a collection of prompts, subscribe to Practicaly AI. We publish workflows, frameworks, and tutorials for people who want to use AI well without spending their week learning AI tools.
How useful was this guide for you?
💌 We’d Love Your Feedback
If you need any guidance while implementing this, or if something isn't quite clear, feel free to ping the team. We're here to support you and clear things up.
Until next time,
Team PracticalyAI
