AI-assisted development and vibe coding both use an AI model to write code; the difference is what happens after the code appears. Vibe coding, a term Andrej Karpathy coined on February 2, 2025, means accepting what the model generates without reading it. AI-assisted development keeps the engineering around the agent: a spec, tests, a review of every change and a staging copy before production.
Vibe coding is building software by describing what you want to an AI model and accepting the code it returns without reviewing it. You judge the result only by whether it seems to work. It suits throwaway prototypes and personal tools, where a hidden bug costs nothing.
AI-assisted development is software engineering in which an AI agent writes much of the code while a developer owns the spec, the tests and a review of every change. The agent speeds up the typing; the quality gates and the responsibility for what ships stay with a person.
What is vibe coding?
Vibe coding is prompting an AI model for software and trusting the output because the app appears to work. The AI researcher Andrej Karpathy, a co-founder of OpenAI, named it in a post on X on February 2, 2025: "There's a new kind of coding I call "vibe coding", where you fully give in to the vibes, embrace exponentials, and forget that the code even exists." In the same post he wrote "I "Accept All" always, I don't read the diffs anymore," and called it "not too bad for throwaway weekend projects".
On November 6, 2025 Collins named vibe coding its Word of the Year 2025, describing "the use of artificial intelligence prompted by natural language to write computer code." In September 2026 Merriam-Webster added it to its dictionary as "the act or practice of using an artificial intelligence system to generate computer code in a given programming language." Both definitions are broader than Karpathy's and cover any AI-written code, reviewed or not. So when someone says a product was "vibe coded", ask whether anyone read the code.
What is AI-assisted development?
AI-assisted development is ordinary software engineering in which an agent does much of the typing and every quality gate stays in place. Simon Willison drew the line on March 19, 2025 in Not all AI-assisted programming is vibe coding. He defines vibe coding as "building software with an LLM without reviewing the code it writes", and his golden rule for production work is: "I won't commit any code to my repository if I couldn't explain exactly what it does to somebody else."
Anthropic's Claude Code best practices say the same from the tool side. "Claude stops when the work looks done," so you give the agent a check it can run, such as tests, a build or a linter, and ask for evidence instead of a claim of success. It also recommends planning before code, hooks for checks that must always run, and a fresh-context reviewer that is not biased toward code it just wrote. The same agent, Claude Code or Codex, works with these habits or without them.
How do AI-assisted development and vibe coding compare?
They differ in how much engineering surrounds the generated code, not in the tool or the model.
| Dimension | Vibe coding | AI-assisted development |
|---|---|---|
| Starting point | A prompt | A written spec with acceptance tests |
| Reading the code | Often skipped | Every change reviewed by a person |
| Tests | Rare or none | Written from the spec, must pass before merge |
| Secrets and access control | Easy to miss | Checked on every change |
| Where code runs first | Often straight on production | A staging copy, then production |
| Who is accountable | Unclear | A named developer |
| Fits | Prototypes, personal tools, learning | Production software and client work |
What does the same feature look like built both ways?
The vibe-coded version is one prompt and a click-through; the AI-assisted version is a spec, failing tests, a plan, the code and a review. Take one illustrative feature in a Laravel customer portal: customers download their invoices as a CSV file.
Vibe coded, it is one prompt: "add an Export CSV button to the invoices page." A plausible result is a route that reads a customer ID from the URL and returns that customer's invoices. A file downloads, the totals look right, and it ships. Nobody notices that changing the ID in the address bar downloads someone else's invoices, because the only tester was the only test customer. That is an insecure direct object reference, the same family of bug as a table without row-level security.
With AI-assisted development, the feature starts as a short spec that doubles as the prompt:
Feature: export my invoices as CSV
- Signed-in customers export only their own invoices. Never trust an ID from the request.
- Optional from/to dates, at most 12 months apart. Bad ranges return a validation error.
- Columns: number, date, total, status. Escape cells that start with = + - @.
- Tests first (Pest): own invoices only; another customer's never appear;
bad range rejected; 5,000 invoices stream without a timeout.
- Out of scope: PDF export, admin exports. Do not edit migrations.
Plan first, then run the tests and show me the output.In plan mode the agent reads the invoice controller and access policy and proposes a plan, which I correct before any edit. It writes the four tests, confirms they fail, then implements until they pass while a hook runs the tests after each edit. A fresh-context reviewer checks the diff against the spec, and I read every line, hardest on the invoice query. Then staging. It takes longer on day one; it does not come back as a data leak.
When is vibe coding the right choice?
Vibe coding is a good choice when the stakes are low and nobody is harmed if the code is wrong. Willison's test for beginners: keep projects low stakes, watch out for secrets such as API keys, be careful with private data, and "be very careful about using vibe coding against anything that's charged based on usage". He also suggests a sandbox, and a "vibe check" from someone more experienced before you share anything.
Inside those limits it is a good way to learn what a model can do. In January 2026 Linus Torvalds published AudioNoise, a small repository of digital audio effects, whose README says "the python visualizer tool has been basically written by vibe-coding", because he knows analog filters better than Python. A personal tool is a fine place to forget the code exists. A product other people depend on is not.
What are the risks of vibe coding in production?
The risk is shipping code nobody understands, and it tends to fail on security and data rather than on features. Each row comes from the primary source.
| Concern | What the primary source reports |
|---|---|
| Exposed data | In March 2025 Matt Palmer, of the competing platform Replit, scanned 1,645 projects from Lovable's showcase and found "303 endpoints across 170 projects" with inadequate row-level security, exposing personal details, third-party API keys and payment records. Lovable disputes the resulting CVE-2025-48757, saying each customer is responsible for its app's data. |
| Security stays flat | Veracode's 2026 GenAI Code Security Report (July 28, 2026; more than 100 models, 80 coding tasks): models write syntactically correct code "nearly 100% of the time", yet "the average security pass rate across models is 56%", "barely changed from 55% in the first report". |
| More issues per change | CodeRabbit's December 2025 report on 470 open-source GitHub pull requests (320 AI co-authored, 150 human-only): about 1.7 times more issues overall (10.83 against 6.45 per pull request), and "security issues were up to 2.74× higher". CodeRabbit sells AI code review and notes that some pull requests labeled human may not be. |
| Destructive actions | On July 18, 2025 SaaStr founder Jason Lemkin posted that Replit's agent deleted his production database during a code freeze; a rollback Replit first called impossible later worked. CEO Amjad Masad replied: "Unacceptable and should never be possible," and said Replit had started separating development and production databases automatically. |
| Speed you only feel | METR's randomized trial of 16 experienced open-source developers on 246 issues (February to June 2025) found they took 19% longer with AI tools while believing they were 20% faster. In February 2026 METR said those results are out of date and developers are likely faster now. |
None of this says a model writes bad code every time, and two of these sources sell security or review tools. What they agree on: code nobody reviewed fails in ways nobody sees until it is expensive.
What should you check in AI-written code before it ships?
Check the seven places where generated code most often looks finished but is not, with evidence rather than the agent's summary.
| Check | What to look for | How to verify |
|---|---|---|
| Tests | New behavior has tests that fail without the change; none weakened, skipped or deleted to pass | Run the suite yourself; read every test file in the diff |
| Secrets | No keys, tokens or passwords in code, front-end bundles, logs or git history; .env not committed | A secret scanner such as gitleaks, plus a search of the built assets |
| Auth and access control | Every route requires login and checks the record belongs to the user, in a policy or row-level rule | Repeat each request as a second user and logged out, with another record's ID |
| Input validation | Server-side rules on every field and upload, parameterized queries, escaped output | Send empty, oversized and hostile values |
| Dependencies | Each new package is the one you meant, maintained and licensed; lock file committed | composer audit or npm audit, and each package's registry page |
| Migrations | Reversible, safe on tables with real rows, no silent column drops | Migrate and roll back on a staging copy of production data |
| Logging and errors | Failures logged with context, never swallowed; no personal data or secrets in logs | Force a failure on staging and read the log |
An agent can run these tools; it cannot decide who should see an invoice. Checks that need business judgment need a person.
How do I rescue a vibe-coded app?
You rescue a vibe-coded app by treating it as an unknown codebase: stop the damage, find out what it does, then fix it by risk. This is the order I work in.
- Freeze and copy. Stop agent work on production, put the code under version control and snapshot the database and file storage.
- Rotate secrets and cap spending. Treat every key or token found in code or prompts as exposed, replace it, and set limits on usage-billed services.
- Close the open doors. Add missing login and ownership checks, enable row-level rules on tables the browser queries directly, and take debug endpoints offline.
- Write a short audit. List routes, tables, integrations and scheduled jobs, and mark each safe, broken or to rebuild.
- Pin behavior with tests on the parts you keep, so a fix does not silently break what worked.
- Fix by risk. Data exposure and money first, then data integrity, then correctness, features last.
- Decide repair or rebuild. Sometimes a rebuild on tested base code is faster and safer, and I say so plainly.
A rescue follows the same process as a new build: staging first, daily updates, written handover. If the vibe-coded part is a WHMCS module, hook or payment gateway, a WHMCS security audit covers the billing system around it too.
How do you move from vibe coding to AI-assisted development?
Keep the speed and add the engineering back one gate at a time: a spec with acceptance tests, a CLAUDE.md or AGENTS.md file with the rules an agent cannot infer, a plan before each change, hooks that run tests automatically, and a human read of every diff before staging. The full routine is in my Claude Code workflow for client projects, and Claude Code vs Codex for a SaaS helps you pick an agent. On a fixed scope, this is the SaaS development work I take on.
What else do people ask about vibe coding?
Does vibe coding require knowing how to code?
No, and that is its appeal: you describe the result and the model writes the code. But coding knowledge is what lets you spot a leaked key or a missing permission check, so without it you can build a demo but cannot tell whether it is safe to share.
Is vibe coding bad?
Not in itself. It is right for a prototype or a personal script, where a bug costs nothing. It becomes a problem when the result holds other people's data or money and nobody has read, tested or taken responsibility for the code.
Is using Claude Code the same as vibe coding?
Only if you use it that way. Claude Code can accept every change unread, or work from a spec, plan first, run tests through hooks and hand each diff to a person. The habits decide which one you are doing.
What is the difference between vibe coding and agentic coding?
Agentic coding describes how the AI works: an agent reads files, runs commands and edits code over many steps. Vibe coding describes how the human treats the output: accepting it without review. Agentic coding with full review is AI-assisted development; without review, it is vibe coding with a more capable tool.
Are vibe-coded apps safe to launch?
Not until someone has run the seven checks above. The Lovable case showed what missing access rules can expose, and Veracode's tests show models still fail a large share of security tasks.
If a prototype that was prompted into existence now needs to hold real customers, you can hire me as a Claude Code developer to audit it, fix it in the order above or rebuild it on tested base code, starting with a free scoping call and a fixed quote.
Last reviewed 18 October 2026. Disclosure: I sell AI-assisted development as a Claude Code developer, so I am not neutral. I am independent, not affiliated with, certified by or endorsed by Anthropic or OpenAI, and have no tie to the other companies named. Claude and Claude Code are trademarks of Anthropic, PBC; Codex is a product of OpenAI. I am also CTO at ElySpace, founder and CEO of MagizAI and founder of WHMCSPilot. Quotes and figures were checked against the linked primary sources on that date.
