“I built it over the weekend. Can we just use it?”
Monday at Sandline, Sara types NG-200, 40 boxes and a discount into her quote calculator. The Northbay Hotels total looks right. Layla asks for a link.
What is vibe coding good for here? Our view: build the prototype, then choose how to own the system. Let's see what changes when others need it.
What is vibe coding?
Vibe coding is building with AI and judging the result by trying it. You describe, try and ask for changes. In the strict sense, you never read the code.
The original post on 2 February 2025 described giving in to the result and forgetting the code. AI-assisted coding is broader. You can still read, test and answer for generated code.
Sara's prototype is a cardboard shop counter. It helps you find where the till should go. Before customers lean on it, check what holds it up.
Reviewed AI-built software can be serious software. Hand-written code can still lack an owner.
Vibe coding vs agentic coding: who checks?
Vibe coding describes how you accept the result. Agentic coding describes a tool doing development steps.
GitHub's agent mode picks files, edits them and runs commands until done. You review changes and, by default, approve terminal commands. Whether you read them is still your choice.
| Approach | Where it helps | The honest catch |
|---|---|---|
| Prompt, try, adjust | Learning which screen or calculation people actually need | A convincing demonstration leaves implementation questions unanswered |
| AI-assisted development with review | Keeping coding assistance while someone checks requirements and implementation | Review, testing and maintenance still need time and competence |
| Configure an existing platform | Building records, forms and approvals from provided components | You must check the platform's fit and own the configuration |
| Commission a custom application | Implementing requirements that warrant dedicated engineering | You need a durable agreement about support, changes and handover |
Sara can start with a quick screen, then change approach.
When is vibe coding good for business?
Try it when little is at stake and it won't last:
- A one-off chart or calculation on sample data tests an idea.
- A personal file-renaming script affects only you if it breaks.
- A mock-up settles what the sales desk wants, faster than a requirements meeting.
- A working first draft helps Karim write the requirements. Everyone agrees it's a picture.
The builder owns it, using sample or their own data. Nobody else gets hurt if it stops. That's the 0–8 band below.
What changes when people rely on it?
For Sara, a wrong sample total means another edit. At month-end, Hana needs the price and discount rule behind real quotes.
Karim needs to change that rule without breaking the next quote. Preferably before midnight.
From Sara's idea to a process Sandline can own
- Sara, Sales Show it — use sample products; check against a hand calculation.
- Hana, Finance Agree the rule — which prices apply? Which discounts need approval?
- Karim, Systems Check the code — calculations, access, dependencies and what happens on failure.
- Layla, Operations Name an owner — who accepts changes and takes over maintenance?
- Sales desk Release it — use the reviewed version with a written quoting fallback.
GitHub's review guide starts with automated tests and static analysis. Check that packages exist and are maintained. Have a person review the logic, too.
Check that tests weren't deleted or skipped to pass. NIST's secure development framework takes the wider view: secure software needs development practices. Running code alone isn't enough.
Run Beezr's Ownership Test
The Ownership Test is Beezr's model, not an industry standard. It helps plan who owns what. It isn't a security audit or permission to deploy.
Score the use you plan, including customers if they'll use it. Don't score the private demo instead.
Answer ten questions, scoring each 0–3. Add them: the range is 0–30, with bands at 9 and 18. Discuss each answer scoring 3 on its own. This is Beezr's model, not an industry standard or a readiness certificate.
Read each row from left to right.
| Question | 0 points | 1 point | 2 points | 3 points |
|---|---|---|---|---|
| If it gives a wrong result for a day, what happens? | Nothing much | Someone is annoyed | Money or customers are affected | Legal, safety or regulatory trouble |
| Whose data does it touch? | Sample or my own | Internal team data | Customer or supplier data | Personal, financial or health data |
| Who uses it? | Just me | My team | Several teams | Customers or partners too |
| Does it change records in other systems? | No | Reads only | Writes to one system | Writes to several systems |
| Do different people need different access? | No | A little | Yes, by role | Yes, down to specific fields |
| Does anyone need to approve what it does? | Never | Occasionally | For some actions | For most actions |
| Will you need to know who changed what, and when? | No | Nice to have | Yes, managers ask | Yes, auditors or customers ask |
| Who maintains it when the builder is busy or gone? | I will, it's mine | A teammate who understands it | Nobody, really | An outside party we'd have to hire |
| How often will its rules change? | Rarely | Every few months | Monthly | Weekly |
| How long will it live? | Days or weeks | A few months | A year or more | It becomes how we work |
| Total | Band | Message |
|---|---|---|
| 0–8 | Prototype away | A vibe-coded tool is a sensible choice here. Keep it simple and enjoy the speed. |
| 9–17 | Prototype, then plan to own it | Great for learning what you need. Before people depend on it, decide who owns it, where the data lives and how changes get approved. |
| 18–30 | Build it as a governed system | This is business software. It needs permissions, approvals, history and someone accountable for it, whatever tool builds it. |
Method: 10 questions × 0–3 points; bands at 9 and 18.
The result flags each 3 as “Watch these”. Nine mild answers can hide one serious concern in the total.
Sandline · example
Same screen, different responsibility
On Saturday, Sara works alone with sample data. It connects to nothing; she'll delete it in a week. She maintains it. All ten answers score 0: total 0, “Prototype away.”
Now Sandline wants quotes for Northbay Hotels and Dunecrest Facilities. Wrong totals affect money (2); it holds customer data (2). Several teams use it (2); it writes to no other system (0).
Access differs by role (2); some discounts need approval (2). Managers want history (2); no maintainer is named (2). Prices change monthly (2); this becomes how Sandline quotes (3).
Total: 19, “Build it as a governed system.” Dependence is the score-3 watch item. Maintenance scored 2, but Karim will feel it first.
The prototype hasn't got worse. Its job has changed.
In Beezr, Quotes and Products would be modules. A relation picks the product; a Smart Script calculates the price. A workflow approval handles discounts that need a decision.
See the data, developer and workflow pages. Smart Script uses JavaScript at ECMAScript 5.1 level: var and function, with no arrow functions.
What should you check before sharing it?
Ask the owner to show a valid quote and an exception. Try a missing product, invalid quantity and a discount they can't approve. Check whether tests were removed to pass.
Ask where the code or setup lives. How do you release changes, restore a known version and get help?
Have the next maintainer make a small rule change. Try the handover.
How Beezr does it
History in Beezr. Enable Track changes on the quote's status lookup. From then on, it records old and new values, who changed them and when. Approval history shows decisions; workflow run history shows every run, steps and errors. See Beezr's security controls.
Choose by what depends on it
- If it's disposable with small consequences → vibe code it — use sample data and an expiry date.
- If people will rely on it → plan ownership first — name who maintains it, holds data and approves changes.
- If money, sensitive data or approvals are involved → check the controls — get qualified help beyond the score.
- If a platform fits → configure and verify it — own the rules and the setup.
Myth / Reality
“AI-built means unsuitable.” The method alone doesn't settle it. Check what it does, who may do it and who maintains it. A working demo doesn't settle those questions either.
Our take
Our take. Sara's weekend taught Sandline what sales wanted. Give someone time and authority before turning that discovery into a lasting duty.
Key takeaways
- If you're technical: check maths, dependencies, access and failures. Plan a handover someone can use.
- If you're not: ask who changes discounts when Sara's away. Have them show you.
- Score the planned use, then discuss each risk.
- Include upkeep in your ownership cost comparison.
Build Sara's calculator in Beezr
Agree the process with the prototype, then build it.
- Create Quotes and Products modules — quantity, discount and product relations.
- Arrange the form — group inputs; make the price a Smart Script smart field.
- Set permission profiles — fields each role can edit, read or not see.
- Add discount approval — route to a person or role, with approve and reject branches.
- Enable history — track the status lookup from enablement; inspect approvals and workflow runs.
Questions readers ask
Is vibe coding bad for business software?
It suits some experiments and personal tools. For shared work, check the code and who owns it. The test helps you talk; it doesn't approve software.
Can non-coders use vibe coding?
Yes, to explore what a tool should do. If you can't inspect the code, find someone who can. Do that before you rely on it for work that matters.
What are the main vibe coding risks?
For Sara: wrong totals, wrong access, unapproved discounts and no maintainer. Use the ten questions to find yours.
Is vibe coding the future of software development?
Nobody knows for sure. AI helps with prototypes and drafts people then review. Judge today's needs and who can maintain the result.
How is vibe coding different from agentic coding?
One describes accepting results; the other describes tools doing development steps. More steps done by an agent still need review and ownership.
Sandline Supplies is a fictional example company.
Sources
- Original vibe coding post · Andrej Karpathy · 2025-02-02
- Review AI-generated code · GitHub · Undated; accessed 2026-10-17
- Using agent mode in your IDE · GitHub · Undated; accessed 2026-10-17
- Secure Software Development Framework, Version 1.1 · NIST · 2022-02-03