Adding a coding agent to the build stage is not a complete lifecycle. A dependable AI-assisted SDLC makes context, handoffs, verification, and human authority explicit from planning through merge.
1. Plan and specify
Define the intended outcome, affected users or systems, constraints, exclusions, acceptance criteria, and verification. Keep the brief collaborative so product decisions and engineering context reach the same artifact.
2. Research uncertainty
When the implementation boundary is unclear, run a read-only investigation first. Map relevant code paths, dependencies, risks, options, and checks. A teammate should accept or revise the recommendation before implementation begins.
3. Implement in isolation
Give Claude or Codex the accepted brief inside an isolated repository sandbox. Preserve the terminal, changed files, diff, and preview so the team can evaluate the work as an engineering artifact rather than a confident chat response.
4. Verify against the brief
Run available tests and repository checks, then use Reviewer, QA, or Security for distinct questions where appropriate. Reports should identify what was tested, what failed, and what remains uncertain.
5. Review, merge, and learn
Inspect the pull request, evidence, unresolved risks, and product trade-offs. A human approves the merge. Feed failures, review burden, and missing acceptance criteria back into the next spec and reusable recipe.