Cubis Engineers

AI in Development

Use AI to improve engineering work without giving up judgment, security, or ownership.

EngineeringFoundationUpdated Aug 13, 2026aisoftware-engineeringdeveloper-experiencegovernance

AI can explain unfamiliar code, draft tests, compare options, and automate bounded work. It can also produce confident errors, insecure code, invented facts, and unnecessary content. The engineer using it remains responsible for the result.

Core rule

Treat AI output as an untrusted proposal. Understand it, verify it, and own it before it becomes part of the product.

Team rules

  1. Use only approved tools for company work.
  2. Never send secrets, credentials, customer data, or restricted company information to an unapproved model.
  3. Give the minimum relevant context; state the goal, constraints, and non-goals.
  4. Keep changes small enough for a human to understand and review.
  5. Verify claims against code, tests, runtime behavior, and authoritative sources.
  6. Apply the same security, quality, accessibility, and review standards as human-written work.
  7. Do not let an agent perform high-impact or irreversible actions without explicit approval.
  8. Record important assumptions and disclose material AI use when policy, licensing, or review requires it.
  9. Stop when the model lacks context or evidence. Do not fill gaps with plausible guesses.
  10. The author and reviewer—not the tool—are accountable for the merged change.

Do and do not

DoDo not
Ask for options, risks, and evidenceAsk for a large feature and merge the first answer
Share narrow, sanitized contextPaste credentials, private data, or an entire repository without need
Read every changed lineApprove a diff because it looks polished
Run relevant tests and security checksTreat generated tests as proof by themselves
Verify packages, APIs, and citationsTrust invented libraries, versions, or links
Keep a human approval step for impactGive broad production access to an autonomous agent
Rewrite vague or repetitive outputPublish low-value generated content to appear complete
Preserve uncertainty in the recordTurn an assumption into a confident statement

Choose the level of control

WorkRequired control
Explanation, brainstorming, test ideasEngineer checks relevance and facts
Code, configuration, migrations, dependenciesEngineer reviews the diff and runs targeted checks
Authentication, payments, security, customer dataDomain review, threat-aware tests, and explicit approval
Production, deletion, privilege, or external communicationHuman approval at the action boundary and a recovery plan

Learning path

References

On this page