You inherit an AI codebase that runs on the previous developer's laptop but cannot be deployed from the repository. Before estimating new features, establish whether the team has the access, dependencies, and operating knowledge to maintain what already exists.
An inheritance review should reduce uncertainty without treating every unfamiliar design as evidence that a rewrite is necessary.
Start with ownership and reproducibility
Ask who controls the repository, deployment, domain, database, model-provider accounts, storage, and billing. Record owners without copying secrets into the review document.
| Review area | Question |
|---|---|
| Build | Can an authorised new environment run documented steps? |
| Data | Which records are production, test, or local-only? |
| Integrations | What external services and permissions are required? |
| Operations | Where do failures and pending work appear? |
| Recovery | What has actually been backed up and tested? |
Do not begin by executing undocumented migration or cleanup commands against production.
Trace one critical user journey
Choose the outcome the business relies on most. Follow input, authorisation, model use, external actions, stored records, and user confirmation.
Mark which components are implemented, simulated, manually operated, or currently broken. A polished screen can conceal a workflow that depends on someone copying data between systems.
For a fictional document assistant, verify source ingestion, access filtering, retrieved evidence, and answer display before judging how much new functionality can be added.
Separate known defects from missing evidence
A failed build is observed evidence. “The architecture will not scale” is a hypothesis until the reviewer defines the workload, constraint, and measurement.
List unresolved questions with the action needed to answer them. That makes the next engineering phase estimable without inventing certainty.
The portfolio-evaluation guide explains why contribution scope and product maturity need distinct labels.
Agree on a stabilisation milestone
A useful first milestone can establish reproducible setup, one complete workflow, controlled test data, and a documented deployment or recovery path. It should produce reviewable evidence rather than a broad claim that the code was “cleaned up.”
Preserve the current working version and relevant data before changes. Review access and credential transitions through the approved account processes, especially when contributors leave.
The technical-handover guide lists material the incoming team should receive.
Codebase-inheritance questions
Should the new engineer immediately choose a different framework?
Only if evidence and delivery needs justify the migration cost.
Does a successful local run prove production readiness?
No. Verify environment, data, external access, and operations separately.
What should the founder request first?
Request an access inventory and one traced critical journey for the review brief.