Pressure to Change (maryrosecook.com)

During a Notion quality-sprint to boost test coverage, an engineer experimented with AI-augmented workflows to go from function discovery to pull request almost autonomously. They combined a Claude “Code skill” (a step-by-step testing process referencing the team’s testing guide, mocking tips, and nearby coverage) with Codex-driven prompts that: read guidance, write tests, create branches, commit, perform a review pass, and open PRs. They also built a “Final Review Before Pushing Straight to Production” prompt that fed the model high-stakes checks plus auto-fetched human PR comments and prior redirect prompts. The workflow mostly worked—dozens of new tests and several successful PRs—though a batched attempt (four functions ➜ four PRs in one run) failed due to linting errors and iteration friction. This is significant because it demonstrates practical, end-to-end AI assistance in software maintenance: prioritization (pointing the model at a doc of critical/complex functions), autonomous test generation, and automated PR iteration from reviewer comments. The experiment highlights key technical caveats—model/tool mismatch (Claude skill used via Codex), CI/linter integration, workspace isolation (limited git worktrees), and the remaining need for human oversight. It shows AI can materially accelerate testing work, but robust toolchain integration and guardrails remain essential for reliable, repeatable automation.