
What an AI-built codebase is missing (the assessment checklist)
The product works, but nobody can say which parts are safe to touch. This read-only assessment lists what an AI-built codebase lacks, with the evidence for each finding, and changes nothing: your product owner decides what happens next.
The short answer
The usual gaps in an AI-built codebase are tests, review discipline and a record of what the model decided.
A read-only assessment checks for those three, plus secrets kept in the code, an unknown dependency list and errors nobody hears about. It writes each gap down as a finding with its evidence and changes nothing.
What goes unseen until the first serious change?
Your product works and people use it. Nobody, you included, can point to the parts that are safe to change, because the model that wrote most of it kept no notes and the person who prompted it has moved on to the next feature. The gaps stay invisible until the first serious change or the first outside review, and then they arrive as a surprise instead of as a list.
Who asks for this inventory?
A founder or technical lead with a working AI-built product who suspects it is fragile and wants a named list of gaps before spending on fixes. It also suits a team about to hand the code to an engineer who was not there when it was built.
What does each finding carry?
The output is a findings report. Every entry in it has the same parts, so a worry becomes something a person can check.
| Part of a finding | What it records | Why it is there | Who uses it |
|---|---|---|---|
| What is absent | A plain statement of the missing piece | So the gap has a name | Your product owner |
| Where it was seen | The file, module or setting where the absence shows | So anyone can go and look | Your engineers |
| Evidence | What was read, and what was missing there | So anyone can check the finding | Anyone reviewing the report |
| Published reference | The public standard that names the same concern, where one exists | So the finding rests on more than one reader's opinion | Your owner and any outside reviewer |
| Proposed order | An optional suggestion, labelled as a proposal | To help the owner sequence, never to bind them | Your product owner |
| Owner's ruling | Accepted or rejected, by name | So nothing moves on a finding nobody accepted | Whoever acts next |
How it works, step by step
Each step below is one thing the assessment looks for. It reads only what you share and writes down what is absent.
Check for tests around what users rely on
Look for a review trail
Look for a record of the model's decisions
Find where secrets live
Take a dependency inventory
Check whether failures reach a person
Human gate: your owner rules on every finding
What does a read-only assessment leave out?
No edits
It writes a findings report and touches no file. The code you hand over is the code you get back.
No probing of outside systems. An external dependency is listed; how it behaves is a separate piece of work.
No scores
A finding says what is absent and where. It does not grade the product or rank it against others.
No certification
For the security items, the OWASP Application Security Verification Standard provides developers with a list of requirements for secure development, and you can read the findings beside it. Reading the findings beside a standard does not mean the product passes it.
Where this fits
This is the front of the MVP to Stable offer: for a product already built with AI that works, mostly, and now needs hardening into something a team can change with confidence. A vibe-coded MVP is AI without governance, and the first three steps above are that diagnosis applied to code. Acting on accepted findings, one reviewed change at a time, is what the hardening page in this series covers. If your question is about a process rather than code, the checklist for whether a workflow is ready for an agent is the one to run, and the library of use cases has the rest.
Related reading
- The governed research brief: the same habit of tracing every statement to its source, applied to research.
- When not to automate: telling apart what preparation fixes from what should stay with a person.
- Readiness checklist: the process-side counterpart to this code-side list.
