Ai Engineering Culture — Claude Code Guide

AI engineering culture appears in actions

AI engineering culture is not a declaration; it is the team’s typical behavior under pressure: how it responds to errors, a broad diff, and confident AI answers. The developer remains responsible for task boundaries, risk, and merge.

The practical cycle is: task → plan and boundaries → small change → checks → human review → documentation → workflow improvement.

Six team principles

Principle	Practical behavior
Human responsibility	A human defines boundaries, risk, and the merge decision
Actual diff	An AI summary does not replace reviewing files, scope, and configuration
Evidence first	Tests, smoke checks, CI, and logs matter more than a confident tone
Root cause	Reproduction and a hypothesis matter more than a patch that hides the symptom
Docs match code	README, runbook, and release notes are checked by actually running the system
Failure changes the process	A recurring problem improves the policy, skill, hook, gate, or template

Review the actual diff and causes

compare the file count with the original scope;
find unrelated refactoring, formatting, and configuration changes;
reproduce the bug with a minimal test or scenario;
confirm that the regression test captures specific behavior;
if AI treats a neighboring symptom several times, return to discovery;
do not claim performance or production readiness without appropriate measurement.

Documentation as part of quality

README, runbook, PR description, and release notes must match the system’s current behavior. Run commands must be executed for real, not copied by analogy from another project.

For changes with recurring consequences, AI_CODING_POLICY.md, CLAUDE.md, REVIEW_CHECKLIST.md, EVIDENCE_LOG.md, SKILL.md, CI logs, and a postmortem are useful. They should explain not only the outcome but also the boundaries and residual risks.

What to do after a repeated failure

Record the symptom and the context in which it recurs.
Find the root cause and separate it from the superficial workaround.
Name the workflow artifact that must change: policy, skill, hook, gate, or PR template.
Assign an owner and add a way to check that the change helped.
Update the documentation and leave a short postmortem.
Check that the new rule did not create excessive false positives.