The Only AI Prompting Guide That Works On Reasoning Models (And Our Cognition)
19 model-agnostic techniques that change how you think, not just what you type. All tested on one prompt. A prompt engineering framework for 2026.
I’ve always wanted to write about prompting, just not in the way you’d expect.
The internet is full of "prompt engineering in 2026" guides and “mega bundles“.
They're generic, break with every model update, and behave differently across models.
With the rise of reasoning models, some techniques can actually hurt the output.
And worse: they turn prompting into mindless copy-paste, not frameworks to understand.
I’ve been testing different prompting techniques since 2022, and spent the last 6 days running one experiment across all of them for this article.
The result is the only prompting framework I've seen that maps how techniques connect, what's happening behind the scenes, and what it means for our cognition.
I drew these techniques from 12 disciplines, not just computer science: foundational prompting, problem reframing, critique, memory, multi-agent, meta-cognition, philosophy (Vedic), science, music theory, physics, economics (game theory), anthropology.
That's 19 techniques from 12 distinct fields applied to one test prompt.
Watch what changes.
Not only in the output, but in your own thinking.
Hey, I’m Karo 🤗
I’m an AI Product Manager and builder. I write Product with Attitude, a newsletter about building with AI and developing critical AI literacy through practice.
If you’re new here, welcome! Here’s what you might have missed:
Claude Cowork Guide for Power Users: 50+ Tested Tips on Plugins, Skills, Sub-Agents, and Memory
Perplexity Computer: What I Built in One Night (Review, Examples, and How It Compares to OpenClaw and Claude)
What's Inside
Which prompting techniques degrade reasoning models and which ones improve them. The Design/Evaluate framework applied to every technique. One running example across 12 tiers. Model-agnostic methods that work on GPT-5.4, Claude 4.6, and Gemini 3.1 without breaking.
What Is Prompt Engineering In 2026?
Prompt engineering in 2026 is about selecting the right cognitive frame and technique for the right problem.
Put simply, prompting is thinking design.
And with the MIT study showing how AI can decrease our cognition, we should probably take every opportunity to design our thinking on purpose.
What Prompting Techniques Hurt Reasoning Models?
By 2026, reasoning models such as OpenAI’s GPT-5.4, Claude 4.6 Opus with extended thinking have internalized chain-of-thought.
They reason internally before producing output, so we don’t have to script the reasoning anymore.
Anthropic takes this further with Claude’s thinking budget that lets you control the depth of reasoning directly, no prompt gymnastics required.
Adding explicit reasoning instructions interferes with a process already running, one built by teams who understand that model's reasoning better than we ever will.
Five Techniques That Degrade Performance On Reasoning Models
Chain-of-thought prompting (Think step by step: first X, then Y.)
Reasoning models already do this internally. Adding an explicit path creates redundancy or even contradiction with their native process.
Few-shot prompting (providing examples to demonstrate structure) PromptHub’s research shows that examples add extra context that overwhelms or redirects the model's internal reasoning instead of guiding it.
Self-consistency prompting (running the same prompt 3 times and comparing)
This technique was built to smooth out the randomness of standard LLMs. Reasoning models are significantly more consistent by design.
Least-to-most prompting (decompose into ordered subproblems) prescribes a reasoning sequence the model handles better on its own.
Skeleton-of-thought prompting (generate structure before content): same problem. It constrains a decomposition process the model would execute more effectively without instruction.
What separates harmful from helpful is clear:
Prescribing reasoning paths hurts performance. Defining goals and constraints improves it.
The rest of this guide is built on that principle.
What Is Model-Agnostic Prompting?
Model-agnostic prompting works across standard LLMs and reasoning models because it operates on goals, constraints, and context, not prescribed reasoning paths.
The techniques in this guide share three properties:
They tell the model what to produce, not how to think
They surface the human’s judgment before the model generates
They keep the evaluation responsibility on the human side
Model architectures change. These techniques don't.
The Two Phases That Change How You Think About Prompting
Every prompting technique operates across two phases:
Prompting without evaluation is consumption. Prompting with evaluation is thinking.
The Experiment
To make this practical and relatable, I picked one simple scenario and ran it through every framework:
Help me draft an About Me section for my website.Then, I compared the outputs and tracked:
how the output changes
how my thinking changes.
Tier 1: Foundational Prompting For Reasoning Models
Zero-Shot Prompting
What Zero-Shot Prompting is
Zero-shot prompting asks a model to complete a task without examples, constraints, or context, relying entirely on its pre-trained knowledge.
Zero-Shot Prompting: About Me Experiment
Help me draft an About Me section for my website.Zero-Shot Prompting: Design Phase and Our Cognition
Zero-shot asks nothing of us in terms of context, preparation or critical thinking. We rely entirely on the model’s pre-trained knowledge:
No framing
No criteria
No constraints
👉 No cognitive effort upfront.
Zero-Shot Prompting: Evaluate Phase and Our Cognition
Not much to evaluate; the output is correct, but generic.
👉 Consumption.
Zero-shot asks nothing of us in terms of context, preparation or critical thinking.
Role Prompting
What Role Prompting is
Role prompting assigns the model a specific identity or perspective, shaping how it interprets and responds to the task.
It’s also the internet’s most over-explained prompting technique. Every $9 prompt guide on Gumroad leads with this.
Role Prompting: About Me Experiment
You are a personal branding strategist who has written 500 About Me sections for consultants and founders. Help me draft an About Me section for my website.Role Prompting: Design Phase and Our Cognition
We can't type on autopilot
We have to make a deliberate decision: which lens do I want this problem solved through?
👉 We’re already thinking about the problem differently than if we’d sent a zero-shot prompt.
Role Prompting: Evaluate Phase and Our Cognition
The model will pull in material about personal branding best practices, positioning for credibility, trust-building copy, and conversion-focused bios.
It will combine that material with anything it already knows about us and produce a strategically crafted About Me section, not a generic bio.
👉 We evaluate the output shaped by one expert's priorities: what that lens included, and what it excluded. Every role is a bias.
Role prompting is the first technique that requires thinking at both ends: choosing the lens before we prompt, and questioning what that lens missed after we read.
Instruction-Based Prompting
What Instruction-Based Prompting is
Instruction-based prompting provides explicit directions for format, tone, constraints, and structure. It tells the model what to produce, not how to think about it.
Instruction-Based: About Me Experiment
Help me draft an About Me section for my website. Follow these instructions exactly:
Format: Return a narrative bio, not a list of credentials.
Scope: Focus only on professional identity — no personal hobbies or filler.
Constraints: No buzzwords like "passionate" or "driven." No third person. No stock phrases.
Tone: Direct, specific, and warm. Skip anything that could appear on anyone else's site.
Length: Under 150 words.
Output: Three sections — Who I am, What I do, Why it matters. Each with a label and a draft.Instruction-Based: Design Phase and Our Cognition
We can’t just describe what we want, we have to prioritize it.
Format or tone? Length or depth? Narrative or structure? Every instruction is a tradeoff we resolve before the model sees the prompt.
👉 We define what matters. And by doing so, we decide what doesn’t.
Instruction-Based: Evaluate Phase and Our Cognition
The model will follow our instructions precisely. That’s the problem.
If we said “under 150 words,” we get 148 words, even if the idea needed 200.
If we said “no buzzwords,” we get clinical phrasing, even where warmth required one.
👉 We evaluate precision against intent. Did the constraints we chose produce what we actually wanted, or did they cut too deep?
Instruction-based prompting forces us to prioritize before we prompt and evaluate whether our own constraints helped or hurt after we read.
Context Stuffing Prompting
What Context Stuffing Prompting is
Context stuffing provides detailed background information so the model can generate outputs tailored to a specific situation.
Context Stuffing: About Me Experiment
Here is everything relevant to my situation:
I'm a product manager with 6 years of experience, currently focused on AI product development
I've been writing a newsletter for 2 years with 500 subscribers. Topics: AI tools, product strategy, building in public
I have no copywriting background but I know what sounds authentic to me.
About Me sections I've liked: [list of examples] and the specific elements I responded to: [details]
My goal: I want inbound consulting inquiries from founders building AI products
My constraint: the About Me needs to work on my homepage AND as a standalone page
What it should NOT be: a resume, a LinkedIn summary, or a timeline of my career
Given all of this, draft an About Me section for my website.Context Stuffing: Design Phase and Our Cognition
We can’t just ask for output, we have to articulate our own reality first.
Who am I? What do I actually want? What have I tried? What doesn’t work?
👉 We’re forced to know our own situation well enough to describe it. That’s harder than it sounds.
Context Stuffing: Evaluate Phase and Our Cognition
The model will skip the generic template and respond to our specific situation.
But context in doesn’t guarantee relevance out. The model may weight the wrong detail, ignore a constraint we buried in paragraph three, or connect two facts in a way that misrepresents us.
👉 We evaluate alignment. Does the output reflect the reality we described, or a version the model found more convenient? And are we sure that the context we provided is accurate?
Most bad outputs are context failures. Context stuffing forces us to articulate our own situation and evaluate whether the model used that reality, or rewrote it.
Tier 2: Problem Reframing Frameworks For Reasoning Models
Step-Back Prompting
What Step-Back Prompting is
Step-back prompting separates the underlying strategic question from the immediate task. We answer the higher-order question ourselves first, before the model ever sees the prompt.
Step-Back Prompt: About Me Experiment
Before we draft an About Me section, answer this first: What is an About Me section actually FOR in 2026? What job does it do? What does a visitor need to believe or feel before they take any action on someone's website?
Once you've answered that, use your answer to draft my About Me section.Step-Back Prompting: Design Phase and Our Cognition
We can’t jump to the task, we have to reframe it first.
Write my About Me is a tactical request. What is an About Me section actually for? is a strategic one.
👉 Step-back forces us to define and ask the strategic question first.
Step-Back Prompting: Evaluate Phase and Our Cognition
The model will reason from principles first, likely something like: An About Me in 2026 is a trust transfer mechanism, not an autobiography
But the strategic frame it chose may not be our strategic frame. It decided the About Me is about trust. Maybe ours is about filtering, attracting the right clients and repelling the wrong ones. Same question, different principle, completely different output.
👉 We judge strategy. Does the model’s higher-order answer match the higher-order answer we would have given? If not, the entire draft is built on someone else’s strategy.
Step-back prompting forces us to rethink the question before we prompt and evaluate whether the model’s strategic frame matches ours after we read.
Tier 3: Critique And Iteration Prompts For Reasoning Models
Self-Refine Prompting
What Self-Refine Prompting is
Self-refine prompting iteratively improves output through cycles of generation, critique, and revision. We define the critique criteria.
Self-Refine Prompt: About Me Experiment
Write an About Me section for my personal website that communicates what I do and why someone should care.
Then critique it: Is it specific? Does it pass the "so what" test? Would a stranger understand in under 5 seconds who I help and how?
Then rewrite it based on your critique.
Repeat this generate-critique-refine cycle two more times. Show all versions.Self-Refine Prompting: Design Phase and Our Cognition
We can’t just ask for a draft. We have to define what makes a draft worth revising.
Is it specific? is a criterion. Does it pass the ‘so what’ test? is a criterion. Each one is a judgment call we make before the model critiques anything.
👉 We define the quality standard.
Self-Refine Prompting: Evaluate Phase and Our Cognition
The model will show three versions, each tighter than the last.
👉 We evaluate progress, not just output. Did each cycle move closer to our standard?
Self-refine prompting forces us to define what "better" means before we prompt and evaluate whether each iteration improved the output or just changed it.
Socratic Prompting
What Socratic Prompting is
Socratic prompting interrogates our question before answering it: defining key terms, surfacing assumptions, and identifying the real underlying question.
Socratic Prompting: About Me Experiment
Before you answer my question, apply the Socratic method to it.
My question: "How do I write an About Me section for my website?"
First: What does "About Me" actually mean? What assumptions does that phrase carry?
Second: What is the question underneath this question? What am I really asking?
Third: What would I need to believe to ask this question the way I'm asking it?
After working through those, answer the real question, not the surface one.Socratic Prompting: Design Phase and Our Cognition
We can’t just ask our question. We have to interrogate it first.
We challenge our own framing.
👉 Socratic prompting forces us to ask: is the surface question even the right one?
Socratic Prompting: Evaluate Phase and Our Cognition
The model deconstructs the surface question and finds the deeper one underneath.
It will likely conclude that how do I write an About Me is actually how do I earn a stranger’s trust in 30 seconds.
But is that the real question? Maybe the real question is how do I filter out clients who aren’t a fit before we ever talk.
👉 We evaluate validity. Did the model find the real question underneath? The redefinition is only useful if it’s ours.
Socratic prompting forces us to interrogate our own question before we prompt and evaluate whether the model’s redefinition matches the question we needed to ask.
Reflection Prompting
What Reflection Prompting is
Reflection prompting asks the model to analyze its own output and identify uncertainty and assumptions after generating.
Reflection Prompting: About Me Experiment
Draft an About Me section for my personal website.
After your draft, reflect on your reasoning: Where were you most uncertain? What assumptions did you make about me that might be wrong? What would change your draft most significantly? What would a critic of your reasoning say?Reflection Prompting: Design Phase and Our Cognition
We can’t just ask for a draft, we have to anticipate where the model will guess and build the reflection questions before we see any output.
👉 We design for transparency.
Reflection Prompting: Evaluate Phase and Our Cognition
The model produces a draft followed by an honest audit of its own reasoning.
The reflection section typically reveals every assumption the model made without evidence.
But self-reported uncertainty is not the same as actual uncertainty. The model may flag surface-level assumptions while missing the deeper ones. It may perform honesty without being honest.
👉 We evaluate confidence. Is the model’s self-assessment trustworthy, or is it telling us what a “reflective model” is supposed to say?
Reflection prompting forces us to design for the model's uncertainty and evaluate whether its self-assessment is honest or performed.
Tier 4: Memory And Context Prompting For Reasoning Models
RAG (Retrieval-Augmented) Prompting
What RAG Prompting is
RAG is the most reliable technique for decisions that depend on current, specific, or proprietary information. By 2026, it's become the enterprise standard for factual accuracy.
It uses external documents as the primary source of truth, grounding outputs in real material rather than model assumptions.
RAG Prompt: About Me Experiment
[DOCUMENTS LOADED]
My last 10 newsletter issues
Three consulting competitors' About Me pages
My most successful LinkedIn posts (top 5 by engagement)
Testimonials from past consulting clients
Based only on these documents, draft an About Me section for my website that differentiates me from competitors and uses my own proven language.RAG Prompting: Design Phase and Our Cognition
We have to curate the evidence the model is allowed to use.
👉 We build the source of truth. The model generates from whatever we load. If the evidence is incomplete or skewed, the output inherits that.
RAG Prompting: Evaluate Phase and Our Cognition
The model draws conclusions grounded in the documents you loaded rather than general training knowledge.
But grounding is not the same as accuracy. The model may cherry-pick the most quotable line while ignoring the most representative one. It may blend language from two documents in a way that sounds like us but says something we never said.
👉 We evaluate grounding. Did the model use our documents as the source of truth, or as a style guide? There’s a difference between based on and inspired by.
RAG prompting forces us to curate what counts as evidence and evaluate whether the model stayed grounded in that evidence or drifted from it.
Tier 5: Multi-Agent Prompting For Reasoning Models
Multi-Agent Prompting
What Multi-Agent Prompting is
Multiple perspectives analyze the same problem simultaneously, then synthesize into a single output.
Multi-Agent Debate produces +11% factual accuracy over single-model answers and is how many agentic systems work internally.
Multi-Agent Prompting: About Me Experiment
Run a debate between two positions:
Agent A argues: "An About Me section should be minimal. Two sentences, one CTA. Anything more dilutes the signal and loses the visitor."
Agent B argues: "An About Me section needs depth. Your story, your philosophy, your proof. Clients need to see how you think, not just what you claim."
Have each agent make their strongest case, then rebut the other's position. After the debate, synthesize the answer that captures the strongest points from both sides.Multi-Agent Prompting: Design Phase and Our Cognition
We have to engineer a conflict worth having.
Which two positions? Why these and not others?
👉 We design the disagreement.
Multi-Agent Prompting: Evaluate Phase and Our Cognition
The model argues both positions at full strength before any synthesis happens.
👉 We evaluate whether the result is a genuine resolution or a false compromise.
Multi-agent prompting forces us to design a disagreement worth having and evaluate whether the synthesis resolved the tension.
You’ve seen 10 of 19 model-agnostic prompting techniques.
Some of the remaining ones include techniques no other prompting guide touches: Philosophy (Vedic), Music Theory (Jazz Prompting), Game theory (Nash Equilibrium Prompting), Anthropology (Thick Description Prompting).
Plus: a pattern analysis across all 19, why 89% of prompting is additive, the unique failure mode each technique carries, and how responsibility shifts from model → human → system.
This isn’t a prompt library. It’s a thinking framework that works on every reasoning model in 2026, and changes how you think.