Taking a vibe-coded app to production means adding back the review and testing that vibe coding skips. Simon Willison's definition is useful here: vibe coding is building software with an LLM without reviewing the code it writes, and his own rule for anything he commits is that he could explain exactly what it does to someone else. Our website was built with heavy help from an AI coding agent: 64 of the 98 commits on its main branch, as of 16 November 2026, carry a co-author line for one. That was the site's first version, which served elrootlabs.com until we moved to a new build on 16 November. What follows are the problems that still needed fixing after a change had run, taken from that commit history, and the check that would have caught each one sooner.
The first build returned 403 on the real host
The first version was a static build. On our shared host it returned 403. Getting it live took three days of commits, and the README now records why: the host starts a Node process only when an entry file is configured, the entry file must be CommonJS because the host loads it with require(), and it must sit at the root directory, or the host reports "Build failed" with a clean log.
None of that is in the code an agent writes, because it is not in the code. It is in the host's behaviour. The check: deploy a hello-world version to the real host on day one, before any features, and write down every setting that made it work. An agent can follow those notes later; it cannot discover them by reading the repository.
The database overrode the code
That first site had a small built-in content manager. Content saved there replaces the arrays shipped in code. So when we later edited an existing project description in code, the build passed, the tests passed and the live site kept showing the old text, because production had a saved copy.
This one is easy to miss because every signal said the change had shipped. The fix was a migration that replaces a stored value only if it still equals the old shipped text, so hand edits made in the content manager are left alone. The check: for any system where data can override code (a CMS, feature flags, a settings table), verify a change on the live site, not in the build output. We wrote that rule into the project's instructions so every later session, human or agent, reads it first.
Styling that passed every test and failed the eye
Our accent colour is a bright lime. On dark backgrounds it works. On the light theme it was being used as text and borders on white cards and as a button fill on a pale navigation bar, where it is close to unreadable. The design document had already warned that lime text on paper fails contrast. The code rendered without errors, and a page-wide colour-inversion filter for the light theme made it look handled.
The fix was an explicit rule: any accent element on a light surface uses black, any accent on a black surface uses lime, set per element rather than approximated by inversion. The same commit caught a callout box that would have turned white text on a white background. The check: look at every page in every theme at phone and desktop widths before merging a visual change. An agent can run that check in a browser if you ask for it; it will not think to unless the instructions say so.
Dependencies: the scan found what the build did not
A security scan from our host flagged 22 advisories in September. All of them were in development and build tooling; the production dependency tree had none. Eighteen were cleared by updates within compatible versions. The remaining four were in one chain whose only fix was a downgrade that would have broken database migrations, so we documented the risk and kept the version.
In November the reverse happened. The stable release of the framework adapter we use imports a module that only exists on another platform, which makes the Node server our host runs fail at startup. We stay on a beta release on purpose, with the reason written in the commit message.
OWASP's 2025 list now has Software Supply Chain Failures as its third category, and Security Misconfiguration as its second. An agent can run npm audit fix in seconds, and nothing in the repository tells it which upgrade breaks your host. The check: run the audit with production dependencies separated from development ones, upgrade within compatible versions first, and after any upgrade start the production server as well as the dev server.
Metadata that was valid and wrong
An SEO review of 16 pages on 1 November found meta descriptions outside a sensible length, titles over 60 characters, and card headings marked as h3 directly under the page's h1. Two weeks later we found every blog post claiming it had been modified before it was published, because the modified date came from the content store's last save rather than the post. All of it was syntactically valid. A validator would have passed every page.
The check here is a test that reads the output the way a search engine does. We added one that walks every URL in the sitemap and checks the title, description and heading order, and another that fails if a post's modified date is earlier than its published date.
The checklist we use now
Most of these problems share one cause: the agent could see the code and could not see the environment the code runs in. The checks that worked are the ones that bring the environment into view:
- Deploy to the real host first and record every setting that made it run.
- List the places where stored data can override code, and verify changes on the live site.
- Review every visual change in each theme and at two widths, in a real browser.
- Separate production from development dependencies in audits, and start the production server after any upgrade.
- Test the rendered output, such as titles, headings, dates and links, from the sitemap.
- Keep a written project file of the lessons above, and have every session read it before it writes code.
- Merge through a pull request a person reads. Our blog posts follow the same rule: an automated gate checks them, and a person merges them.
For the trade-off between building something like this and buying a ready-made product, see custom software versus off-the-shelf. For the structured-data side of the same site, see our schema markup audit.
