So, Are You Really Vibe Coding? The 2026 Definition, Lexicon, and Decision Framework
AI building became a spectrum in 2026. Here's the updated definition, the A–Z lexicon many builders keep getting wrong, and how to choose the right build mode before it touches real users.
TL;DR: Vibe coding is using natural language prompts to model, prototype, and build products with AI: you describe what you want in conversation and iterate on what the AI generates instead of writing every line yourself. In 2026 it is no longer one workflow. It is a spectrum, from fast conversational prototyping to spec-first AI-assisted development, and your job is knowing how much specification a project needs before it touches real users. The term was popularized by Andrej Karpathy on February 2, 2025, and on February 8, 2026, almost exactly one year later, he reportedly called the framing passé and moved past it toward agentic engineering. This guide gives you the updated 2026 definition, a full AI development lexicon (product thinking, design thinking, prompt engineering, context engineering, vibe coding, AI-assisted engineering, spec-driven development, agentic coding, loop engineering, agentic engineering, and critical AI literacy), and a 5-question decision framework for choosing how much structure your build needs.
How sure are you that what you’re doing is vibe coding?
I’m asking because many people use this term loosely and treat it as a synonym for any AI-assisted coding.
Some people, like me, use the term loosely on purpose.
Others are simply not yet familiar with the distinction.
Read this guide even if you’re already a pro builder.
Use it as a reference. Print it if you still believe in paper. Send it to your team if they keep calling everything “vibe coding.”
It gives you the vocabulary to describe what kind of AI builder you are.
What’s Inside
Hey, I’m Karo Zieminski 🤗
AI Product Manager and builder.
I write Product with Attitude, an AI newsletter for thousands of subscribers developing critical AI literacy the only way it sticks: through practice.
We don’t just use AI. We build workflows, automations, and products with it, while studying how AI itself is built, positioned, and woven into our work.
If you’re new here, welcome! Here’s what you might have missed:
I Built a Runaway Subscribe Button. It Brought 70 New Subscribers in 6 Days.
I Built A Tool Every Substack Writer Needs. And It’s Not AI.
Part 1:
The 2026 Definition of Vibe Coding
Vibe coding is using natural language prompts to model, prototype, and build products with AI. You describe the product in conversation. The AI generates the code. You react to what it built, refine the prompt, and iterate until it works, or until you understand why it doesn’t.
Three things separate vibe coding from AI-assisted coding:
1. Natural language is the primary interface. You’re steering with intent, not syntax.
2. You evaluate outcomes, not code. You judge the running product, the behavior, the failure modes, without reading every line the AI wrote.
3. Iteration replaces specification. The plan takes shape through the conversation instead of being written upfront.
The term comes from Andrej Karpathy, who described the practice on February 2, 2025.
What started as a meme about “giving in to the vibes” became the default way thousands of non-traditional builders build real software.
By early 2026, Karpathy himself had moved on.
On February 8, 2026, he reportedly called the original framing passé and redirected attention toward agentic engineering, as covered by kingy.ai.
Vibe coding raises the floor, so non-technical people can build.
Agentic engineering raises the ceiling, so expert builders can ship far more.
That marks where the vibe coding term sits today: one mode on a wider spectrum. More on that below.
What Vibe Coding Is Not
The term gets messy because people keep merging four separate things.
Vibe coding is not a synonym for “any AI-assisted coding.”
This is probably the one making the term blur fastest.
Simon Willison called it semantic diffusion: a precise term spreads, people use it loosely, and the original meaning gets muddy.
By 2026, the split is obvious.
In casual conversation, “vibe coding” often means any kind of building with AI.
But in more serious technical writing, the meaning has narrowed. It describes the low-governance end of AI-assisted building: shipping without enough review, structure, or accountability.
A senior engineer who uses AI to draft a function, then reviews every diff, is doing AI-assisted engineering.
A beginner who chats an app into existence and discovers authentication later is vibe coding.
Same tools.
Very different accountability.
Vibe coding is not no-code.
No-code gives you pre-built blocks. Vibe coding produces real code you own, can read (or have AI explain), and can take anywhere.
You notice the difference when you outgrow the platform, or when the platform changes and your compliance risk changes with it.
Vibe coding is not slop coding.
Slop coding is vibe coding minus judgment, shipped anyway.
Speed without judgment produces AI slop: generic, emotionless, low-quality output generated without human taste, or intentional direction.
Slop coding is what happens when a poorly executed attempt at vibe coding turns your build into a magnificent disaster: you outsource your taste, compromise on product thinking, or skip intentional direction.
Vibe coding is not the enemy of good engineering.
Vibe coding and spec-driven are two ends of the AI-assisted software development spectrum. In reality, many serious builds live somewhere in the middle: fast enough to learn, structured enough to survive contact with users.
Some projects can start with a rough idea, and that’s fine: a prototype, a landing page, a small internal tool, a personal automation, or a feature you are still feeling your way through. Vibe coding gives you a fast iteration loop before the shape of the build is fully known.
Other projects need proper specs from the beginning: user stories, roles, data models, security constraints, compliance requirements, acceptance criteria, edge cases, deployment plans, and migration paths.
Part 2:
The 2026 AI Development Lexicon
Product thinking
Product thinking means understanding users, problems, behaviour, constraints, business value, and product strategy well enough to make better decisions about what to build, why it should exist, and how it should improve over time.
Popularised by
Julie Zhuo, June 15, 2021
What AI does
AI helps map user needs, analyse feedback, compare trade-offs, generate product options, simulate edge cases, and expose gaps in the product logic.
What you do
You decide which problem matters, who it matters for, what outcome the product should create, what trade-offs are acceptable, and whether the solution deserves to ship.
Typical workflow
User → Problem → Outcome → Options → Trade-offs → Decision → Measure → Iterate.
How to think about it
You treat product thinking like judgment under constraint: AI can generate more options, but you still own the decision about value, timing, risk, quality, and consequences.
More on product thinking
Design thinking
Design thinking means using human-centred research, problem framing, ideation, prototyping, and testing to create products, services, and systems that balance user needs, technical feasibility, and business viability.
Popularised by
David Kelley, Tim Brown, and Jane Fulton Suri, 1990s
What AI does
AI helps synthesize research, generate solution directions, stress-test assumptions, create prototypes, and compare design options against user needs and constraints.
What you do
You investigate the real problem, define the user need, shape the design challenge, choose what to prototype, test with people, and decide what should change.
Typical workflow
Research → Define → Ideate → Prototype → Test → Learn → Refine.
How to think about it
Treat design thinking like a human-centred problem-solving loop: AI can speed up the work, but people, evidence, and lived context decide whether the solution is worth building.
Prompt engineering
Prompt engineering means designing the instruction given to an AI system, including phrasing, structure, examples, constraints, and success criteria, so the model produces more useful and reliable output. It’s the skill of asking with enough precision to get useful output.
Popularized by
Lilian Weng, March 15, 2023
What AI does
AI acts as a responsive system that changes its output based on how clearly the task is framed.
What you do
You shape the request, add examples, define the format, set boundaries, and tell the model what good output looks like.
Typical workflow
Goal → Instruction → Constraints → Examples → Output → Evaluate → Refine.
How to think about it
You treat the prompt like the steering wheel: the better the instruction, the less time you spend cleaning up vague, generic, or misaligned output.
Read the full 2026 guide
Context engineering
Context engineering means curating structured, multi-source context packages that give AI everything it needs to perform accurately: documentation, personas, brand guidelines, test cases.
Popularised by
Tobi Lütke and Andrej Karpathy, June 27, 2025
What AI does
AI acts as a reasoning system that depends heavily on the quality, relevance, and structure of the information it receives.
What you do
You gather the right materials, remove noise, organise the facts, and give the model enough context to make better decisions.
Typical workflow
Goal → Gather Context → Structure Context → Remove Noise → Prompt → Evaluate → Update Context.
How to think about it
Treat context like the working environment: the model performs better when it has the right documents, rules, examples, constraints, and memory around the task.
Read the full 2026 guide
Vibe coding
Vibe coding, also called vibecoding, means using natural language prompts to prototype, model, and build software with AI through a conversational workflow where the product is judged by outcomes rather than a traditional upfront specification.
Popularized by
Andrej Karpathy, February 2, 2025.
What AI does
AI acts as a fast builder, interpreter, and improvising technical partner.
What you do
You describe what you want, test what appears, refine through conversation, and keep moving until the product works well enough for the current goal.
Typical workflow
Prompt → Build → Test → Refine → Repeat
How to think about it
You treat the AI like a build partner that can turn intent into working software, but the risk increases when you skip review, architecture, testing, or ownership.
AI-Assisted Engineering
AI-assisted engineering, also called augmented coding, means using AI as a coding collaborator while the human engineer remains responsible for architecture, code review, testing, security, and final approval before anything ships.
Popularized by
Kent Beck, May 03, 2025.
What AI does
AI acts as a collaborator, not a replacement for engineering principles.
What you do
You stay in control, review every line, and maintain tests while the AI accelerates drafting, refactoring, debugging, and repetitive implementation work.
Typical workflow
Task → AI Draft → Human Review → Test → Refactor → Ship.
How to think about it
You treat the AI like a power tool, not an autopilot: the model drafts, you review and refine, and nothing ships without human sign-off.
Spec-driven development (speccoding)
Spec-driven development, also called speccoding, means turning a written product or software specification into executable AI-assisted implementation work where the spec, not the chat conversation, is the source of truth.
Popularized by
GitHub, September 2, 2025.
What AI does
AI acts as an implementation partner that follows a defined plan instead of inventing the product through conversation. The spec, not the conversation, is the source of truth.
What you do
You define what and how should be built, write the rules, create the plan, break the work into tasks, then use AI to implement and test against the specification.
Typical workflow
Prompt → Spec → Design → Tasks → Code → Test Against Spec → Refine → Repeat.
How to think about it
Treat the spec like the contract: the model can help build, but the written requirements decide whether the work is correct.
Agentic coding
Agentic coding means delegating multi-step software development work to AI agents that can plan, edit files, run checks, interpret errors, revise their approach, and report progress with limited turn-by-turn supervision.
Popularised by
Addy Osmani, January 28, 2026.
What AI does
AI acts as a semi-autonomous builder that can move through a task sequence instead of waiting for every next instruction.
What you do
You give the agent a larger objective, such as building a feature, writing tests, fixing deployment issues, or investigating a bug.
You let the AI agent plan, execute, test, and self-correct with minimal turn-by-turn supervision.
The agent then breaks that goal into smaller steps, chooses the sequence of work, edits files, runs checks, reads errors, adjusts the plan, and reports back.
Typical workflow
Objective → Plan → Execute → Check → Fix → Report → Human Review.
How to think about it
You treat the agent like a junior developer with tools: it can do real work, but it still needs scope, guardrails, review, and a clear definition of done.
Loop engineering
Loop engineering means designing the system that prompts the agent, instead of typing every prompt yourself. It’s system design: a repeatable process that creates objectives, assigns them to agents, evaluates the output, and feeds the result back into the next step.
Popularised by
Addy Osmani, June 7, 2026
What AI does
AI acts as a repeatable worker inside a designed loop, moving through tasks, checks, fixes, and updates until the goal is complete or blocked.
What you do
You define a repeating goal once, give the agent a way to find work, act, verify the result, and remember what it did. After that, the system keeps driving the agent until the goal is either met or blocked.
Typical workflow
Goal → Trigger → Find Work → Execute → Verify → Record → Repeat.
How to think about it
Treat the loop like an operating system around your agents: the agent does the work, but the loop decides when work starts, what gets picked up, how progress is checked, and when a human needs to step in.
I don’t prompt Claude anymore. I have loops running that prompt Claude and figuring out what to do. My job is to write loops.
Agentic engineering
Agentic engineering means using AI agents inside a disciplined software engineering process where agents can plan, write, test, and revise code, but humans remain responsible for architecture, quality, security, verification, and the final shipped system.
Popularised by
Andrej Karpathy, February 2026
What AI does
AI acts as a capable but fallible engineering agent that can handle implementation, testing, debugging, refactoring, and iteration across a codebase.
What you do
You define the architecture, write or approve the spec, assign scoped work to agents, review diffs, validate tests, check security, and decide what is safe to merge.
Typical workflow
Spec → Plan → Delegate → Agent Builds → Test → Review → Refine → Merge.
How to think about it
You treat agents like fast junior engineers with tools: they can accelerate serious development, but they do not remove your responsibility for judgment, correctness, maintainability, and production risk.
Critical AI literacy
Critical AI literacy is the ability to understand, use, evaluate, and challenge AI systems with informed judgment.
It combines practical AI fluency with systems awareness: how AI models are built, what data and labour shape them, what incentives drive their design, what tradeoffs they make, who they serve, who they exclude, and what happens when humans outsource thinking, creativity, decisions, or responsibility to machines.
Popularized by
Dr Sam Illingworth and Karo (Product with Attitude).
It’s one of the reasons I started Product with Attitude. I don’t want people becoming gullible users of powerful AI tools. I want them to approach these products with attitude.
What AI does
AI acts as the system under examination: it produces fluent, confident output that can be right, wrong, biased, or shaped by incentives we can't see.
What you do
You question the source, check the claims, notice what's missing, monitor the model’s reasoning, name the trade-offs, and decide when to trust the machine and when to override it.
Typical workflow
Think → Contextualise → Interrogate → Verify → Decide → Stay Accountable.
How to think about it
Treat critical AI literacy like the crucial meta-skill underneath every other entry on this list.
Part 3:
How to Choose the Right Build Mode
Choose the right AI build mode project by project, based on specification: how much must be defined before AI touches live work.
Run every build through five questions before you start.
1. Who gets hurt if this breaks?
If the answer is “just me,” you can stay further left.
Examples: A personal automation. A silly prototype. A birthday website. A tiny tool for yourself. If it breaks, you sigh dramatically, fix it, and move on with your life.But if the answer is users, customers, students, clients, colleagues, patients, employees, or anyone who depends on the thing working, you move right.
The moment other people rely on the product, the build needs more structure: clearer requirements, stronger testing, acceptance criteria, documentation, and a plan for what happens when something fails.
2. Can you state the acceptance criteria out loud right now?
Acceptance criteria means: what must be true for this feature to count as done?
If you cannot say that yet, you’re still in discovery and that’s ok. Vibe code to learn the shape. Once the shape is clear, switch modes and write the spec before you scale it.
This is where AI is useful as a sparring partner. So before you continue writing code, make it interrogate the idea.
If you’re not sure how, use my Pre-Build Interrogation Prompt.
3. Will you touch this code again in three months?
If the answer is no, vibe freely. Some builds are meant to be disposable. You make the thing, learn from it, use it once, and let it go.
But if you expect to return to the code later, you need to move to spec-driven development.
Future-you will not remember why the database table is named like that. Future-you will not remember which files should never be touched.
Trust me.
The technical-debt wall is real: vibe-coded builds hit a documented three-month point where maintenance overhead stacks up.
4. Does a user, regulator, platform, or enterprise customer care how this was built?
If yes, you must document your work. No exceptions.
Regulated products, enterprise workflows, payment systems, safety-critical tools, compliance-heavy products, multi-agent systems, and anything with audit requirements need a spec-first contract mode.
This stops being theoretical on August 2, 2026, when the EU AI Act’s next enforcement phase lands. An undocumented, vibe-coded build history is exactly what regulators are learning to ask about. If you ship to EU users, the audit trail is no longer optional.
Learn more:
5. Are you using AI to create code, product logic, data logic, or decisions you cannot personally explain?
If yes, move right on the spectrum.
You don’t need to understand every line like a senior engineer. But you do need to understand enough to explain what the system does, where it can fail, and who is affected when it does.
That’s where critical AI literacy enters the build process. And where the evidence gets uncomfortable. A 2025 randomized trial from METR found experienced open-source developers were 19% slower when using AI assistance, while believing they were about 20% faster. The risk is not just code you can't explain. It's confidence you didn't earn.
Choosing the wrong build mode is exactly the moment most “vibe coding gone wrong” stories are born: someone kept vibe coding past the point where the build needed a spec.
So, Are You Really Vibe Coding?
If you’re prototyping something only you will see, testing it by whether it runs, and letting the plan take shape through chat, yes, that’s vibe coding. And it is exactly the right mode.
If the build touches real users, real data, or real money, specs, product thinking, design thinking and critical AI literacy stop being optional.
So before your next build, ask yourself the five questions listed above.
They separate a useful product from a future postmortem.
FAQ
What is vibe coding in simple terms?
Building software by describing what you want to an AI in plain language and iterating on what it generates, instead of writing the code yourself.
Who coined the term vibe coding?
Andrej Karpathy popularized it on February 2, 2025. It moved from meme to methodology within months. By early 2026 he had already pointed past it toward agentic engineering.
Is vibe coding the same as AI-assisted coding?
No. Vibe coding is the low-governance end of building with AI: you steer in natural language, judge outcomes over code, and let the plan emerge through conversation. AI-assisted engineering keeps a human reviewing every diff and owning architecture, testing, and security. Same tools, very different accountability.
Is vibe coding the same as no-code?
No. No-code assembles pre-built blocks inside a platform. Vibe coding produces real code you own, can read or have AI explain, and can take anywhere.
Is vibe coding the same as spec-driven development?
No. They’re two ends of one spectrum. Vibe coding lets the plan emerge through conversation. Spec-driven development writes the plan first and holds the AI to it. Most real projects sit in between.
Do I need to know how to code to vibe code?
No. Many creators are already building useful software without traditional coding backgrounds. You need product judgment: knowing what to build, recognizing when output is wrong, and asking the right next question. That’s critical AI literacy, and you build it by building.
What is loop engineering?
Loop engineering is designing the system that prompts an AI agent on a schedule and against a goal, instead of typing each prompt yourself. The term was popularized in June 2026 by Addy Osmani. It sits one layer above agentic coding: you build the loop, the loop drives the agent.
What’s the difference between prompt engineering and context engineering?
Prompt engineering crafts the instruction. Context engineering curates everything the AI knows when it executes the instruction: docs, examples, constraints, state. In 2026, context engineering is where the advantage is.
What is spec-driven development?
Spec-driven development is an AI-assisted building workflow where the specification becomes the source of truth. Instead of asking AI to start coding from a loose idea, you define what the product should do, how it should behave, what constraints matter, and how success will be tested before implementation begins.
When should I stop vibe coding and write a spec?
When other people depend on the thing working.
Also when the build touches payments, permissions, user data, security, compliance, enterprise workflows, education, legal, health, employment, housing, or anything where failure creates real harm. The rule is simple: as stakes increase, structure must increase.
What replaced vibe coding in 2026?
Nothing replaced it — it became one mode on a spectrum. Vibe coding raises the floor for non-technical builders. Agentic engineering and loop engineering raise the ceiling for expert builders shipping production systems. You pick the mode based on blast radius and how much specification the project needs.
Further Reading
WHY SUBSCRIBE ・YOUR BENEFITS・ TOOLS I BUILT・CLAUDE HUB・PERPLEXITY HUB ・VIBE CODING HUB
If I could give this article 5 hearts I would. Great job Karo.
On Product Thinking I am very much a Pragmatic Framework guy and it overlaps with Julie Zhou’s article. My particular emphasis is the upstream and downstream of the development cycle in the product cycle.
Thanks for sharing this compilation of terms Karo!. While development is more accessible than ever, quality software still depends on sound engineering practices. Vibe coding, AI-assisted development, spec-driven development, and 'traditional' development aren't competing.. Knowing which to use and when to use it is what leads to better decisions.