The login that only works on slide 4
The dazzle OAuth in twelve minutes. Investors nodded.
The death The callback was hardcoded to localhost. Production returned 404. Nobody had a password-reset path either.
The methodThree modes of AI coding
The Vibe Coding Curse
It feels like magic. It ships like a demo.
Vibe Coding is the most intoxicating way a person has ever produced software. A prompt, a flourish, something that looks like a product in twelve minutes. The room applauds. Then someone asks about password reset, tenant isolation, or the carbon cost of the polling loop — and the trick collapses.
This page names the curse, names the engineer who beats it (the Code Pilot), and shows why the target operating model is not a better prompt. It is a crew of software engineers running the AI Delivery Lifecycle at the heart of AI Product Studio.
- Modes of AI coding
- 3
- Vibe-coded apps that feel done
- Most
- That actually reach production
- Almost none
- People who still decide what ships
- SEs
Three modes of AI coding
The same models sit under all three. The difference is who is flying them, what they pay attention to, and whether anyone else can land the thing on Monday.
| Mode | Experience | Pros | Cons |
|---|---|---|---|
| Vibe CodingThe Vibe Coder | Low to none as a software engineer. High confidence in the prompt. The tool is the skill. |
|
|
| Code PilotingThe Code Pilot | Deep software-engineering craft. Years of production scars. Agents are a power tool, not a personality. |
|
|
| AI Delivery LifecycleThe Studio Crew | A crew of software engineers, plus product, security and compliance, sharing one operating model. |
|
|
Where the hype scores — and where production lives
Vibe Coding wins the first two rows. Those are the rows a conference audience can see. Production is the other six. Colour is paired with the mode name; the numbers sit beside the bars.
- Vibe Coding
- Code Piloting
- AI Delivery Lifecycle
| Concern | Vibe Coding | Code Piloting | AI Delivery Lifecycle |
|---|---|---|---|
| Demo dazzle | 96 | 72 | 68 |
| Time-to-wow | 94 | 70 | 62 |
| Reliability | 18 | 78 | 92 |
| Security | 12 | 80 | 94 |
| Scalability | 14 | 64 | 90 |
| Green / carbon | 16 | 70 | 86 |
| Auditability | 8 | 48 | 96 |
| Chance of production | 11 | 74 | 93 |
Vibe Coding is a person without strong software-engineering experience using generative tools as if the prompt were the profession. It is not stupid. It is not useless. It is a sugar rush with a product-shaped wrapper.
What it is
The Vibe Coder sits down with an idea and an agent. They describe the feeling of the thing they want. The model produces screens, copy, and a repository that runs. That is a genuine new capability, and it is why the phenomenon has a name, a fandom, and a conference circuit.
Why it is so rewarding
Every accepted prompt is a visible win. A new page. A new button. A dark mode that took four minutes. The brain is not measuring tenant isolation or the absence of a test; it is measuring pixels that were not there ten minutes ago. That is a dopamine loop, and it is honest about what it rewards: novelty, not survival.
Why it is brilliant for demos
A pitch deck needs a working-looking app more than it needs a threat model. A stakeholder workshop needs something to click. A Friday spike needs an answer to “is this even a product?” Vibe Coding is unmatched at those jobs. The curse begins when the demo is mistaken for the delivery.
Prompt
“Build me a dashboard.” The model obliges in seconds.
Dazzle
Charts appear. The room applauds. The hit lands.
Next feature
The unfinished auth flow is already boring. On to dark mode.
Rewrite
A new prompt overwrites last week’s cart. Nothing was tested.
The cliff
Someone asks about tenants, password reset, or the bill. Silence.
The curse is not that Vibe Coding is fake. The curse is that it is real enough to be believed. The dashboard renders. The button clicks. The audience has no reason to ask whether the write path has a condition, whether the secret is in the browser, or whether the “live” chart is a fixture. Hype is what happens when the visible 20% is taken as proof of the invisible 80%.
The demo graveyard
These are not hypotheticals. They are the usual way a vibe-coded app dies on the walk from the stage to production — the less visual work that reliable, secure, scalable, green software actually requires.
The API key in the browser
The dazzle “Test Connection” talked to the model live on stage.
The death The key sat in the client. The first crawler emptied the month’s inference budget before lunch on Monday.
The dashboard of dreams
The dazzle Five charts, three themes, a pulsing KPI.
The death Every series was a fixture in a JSON file. Real telemetry was never wired. The impressive number was a string.
The multi-tenant accident
The dazzle Two customer logos in the header. Instant “enterprise”.
The death One table, no partition key, no field-level access. Customer A opened Customer B’s invoices on the first shared demo.
The greenfield that cooked the planet
The dazzle Live updates everywhere. It felt instant.
The death An unthrottled one-second poll, no cache, a 4 MB bundle of unused UI kits. The carbon cost was never measured because nobody asked.
The “we’ll add tests later” cliff
The dazzle Saturday night: a new cart in one sitting.
The death Monday’s prompt rewrite of the cart broke checkout. There was no suite. There was no owner. There was no rollback.
The WCAG surprise
The dazzle It looked fine on the presenter’s laptop.
The death Status was colour alone. Primary actions were 18-pixel icons. Keyboard users never reached submit. Procurement sent it back.
The GDPR toast
The dazzle A cookie banner pasted from a gist. Job done.
The death No lawful basis, no record of processing, no deletion path. Legal stopped the launch two days before the press release.
The silent 500
The dazzle The happy path never blinked.
The death Empty catch blocks swallowed every write failure. Ops had no logs. Users had no toast. The data simply vanished.
The 2 a.m. last-write-wins
The dazzle Two people, two chats, one file, double the features.
The death No If-Match. No merge. Tuesday morning the schema was whichever prompt finished last. Nobody could reconstruct Friday.
The password-reset ghost
The dazzle Sign-up worked. The first five users cheered.
The death Reset email was a stub that logged to the console. Support opened on day two and never closed.
The weekend schema
The dazzle A whole domain model from a single prompt.
The death Monday: “who owns this?” The fields had no tenant id, dates were Date objects, and the bidirectional refs pointed one way.
The Code Pilot is an experienced software engineer who uses agentic coding tools the way a pilot uses an aircraft: as a machine of extraordinary power that still requires a human who can fly, who can refuse a bad approach, and who is accountable for the landing.
This person is not a slower Vibe Coder. They are the reason software engineering still sits at the centre of the future. They have production scars. They have watched a missing index take down a Friday release. They have rotated a leaked key. They know that “it runs” is the start of the job, not the end of it. Agents make them faster. Agents do not make them optional.
A Code Pilot will throw away a dazzling page that stores a secret in localStorage. They will demand a failing test before the implementation. They will ask who is on call, what the carbon looks like at 10× traffic, and whether a second engineer can reconstruct the decision from the commit trail. That is attention to detail. It is the opposite of the dopamine hit.
Refuse the sloppy win
A feature that cannot survive Monday morning does not ship, however pretty the screenshot.
Review as the job
The Code Pilot reads the diff the agent wrote. If they cannot explain it, it does not merge.
Non-functionals first
Security, reliability, accessibility, scale and carbon are not a later phase. They are the work.
Secrets stay server-side
Keys never touch the browser. Test Connection and Fetch Models go through a backend that already holds the secret.
Tests are not optional garnish
If the agent cannot produce a failing test first, the Code Pilot writes one before any implementation lands.
Leave a trail a stranger can fly
ADRs, threat notes, and a commit message that will still make sense in six months.
Code Piloting still does not scale past one human. The Code Pilot is the correct unit of skill. They are not the correct unit of an organisation. Context dies in their editor. Review is their evening. Evidence is a slide they assemble after the fact. For a team, and for an enterprise, you need an operating model that puts many software engineers at the same gates, with the same memory.
A team of software engineers, plus the people who hold product, security and compliance, running a named agent roster through eight phases. This is the AIDLC — the method at the heart of AI Product Studio. Software engineers do not leave the building. They move to the gates.
Intent
A software engineer accepts or rejects the classified intent.
Context
Constraints, ADRs and NFRs are loaded so the roster cannot invent the architecture.
Plan
The crew approves the work package before a line of implementation is written.
Code
Agents implement against the approved plan. Every change arrives as a reviewable PR.
Review
Engineers answer the agents' questions on each pull request and approve the Definition of Done — on the record.
Verify
Suites, security and compliance evidence must go green, or a human holds the gate.
Release
Two-key sign-off: automation plus a named engineer, every time.
Telemetry
Instrumentation and cost-per-increment feed the next intent. Learning does not reset.
The Vibe Coder asks the model for a feature. The Code Pilot asks the model for a change they are prepared to sign. The Studio Crew asks the model for a run: an intent, a plan, a reviewable implementation, a recorded verdict, a green suite, a two-key release, and telemetry that feeds the next intent. That is not a better chat. That is a lifecycle.
Every unglamorous thing the Vibe Coder skipped — tenant isolation, field-level access, optimistic concurrency, append-only audit, secrets encrypted server-side, accessibility, carbon, DPIA evidence — is a phase with a named agent and a human gate. The dopamine of a new screen is still there. It is no longer the management system.
This is why AI Product Studio exists. Not to make Vibe Coders faster, and not to replace software engineers with a roster. To give Code Pilots a crew, a memory, and a set of gates so the work that never looks good on a conference slide still happens — every run, on the record. Read the AIDLC model and the ROI case.
How the three modes map to the real world
A lone developer, a team, and an enterprise can all touch the same models. They cannot all survive on the same operating mode.
Lone developer
Vibe Coding
A Vibe Coder can produce a dazzling personal prototype in an afternoon. That is a legitimate use. It is not a product.
Code Pilot
A Code Pilot can ship a small production system — with tests, secrets in the right place, and an honest on-call story. Still one thread.
AIDLC crew
Studio Free gives one engineer the eight-phase loop and the audit trail. The Code Pilot stays at the gates; the roster does the rest.
A team of developers
Vibe Coding
A room of Vibe Coders is a room of colliding miracles. Shared understanding lasts until the next rewrite. Reviews become archaeology.
Code Pilot
Several Code Pilots without a lifecycle are brilliant soloists. Context dies in each editor. Quality depends on who had the laptop that day.
AIDLC crew
The Studio Crew is the point. Shared Context Library, gated work packages, named agents, recorded decisions. Software engineers remain front and centre.
Enterprise
Vibe Coding
Vibe coding at enterprise scale is a shadow-IT incident. Ungoverned model calls, unattributable changes, no evidence for the risk function.
Code Pilot
A handful of Code Pilots cannot satisfy a portfolio, an auditor, or a DPIA. Heroics do not industrialise.
AIDLC crew
The AI Delivery Lifecycle is the operating model: policy-bounded agents, cost ceilings, immutable audit, earned autonomy. This is AI Product Studio.
Keep the spark. Lose the curse.
Vibe Coding is allowed to be fun. Prototypes should be cheap. Demos should dazzle. The mistake is promoting the prototype to a product because the room clapped. Software engineers — Code Pilots, then a Studio Crew — are the people who decide what is allowed to live past the applause.
For a spike or a stage
Vibe Coding is a legitimate tool. Name it. Time-box it. Do not let it become the production branch.
For a lone engineer who cares
Fly as a Code Pilot. Refuse the sloppy win. Start a studio on the Free plan and keep yourself at the gates.
For a team or an enterprise
Adopt the AI Delivery Lifecycle. That is the operating model. That is AI Product Studio.
Put software engineers back at the centre
The AIDLC does not ask your Code Pilots to type more. It asks them to judge more — with a roster, a memory, and an audit trail that will still stand up when the demo is a year old.