Ten durable guidelines for agentic coding with tools like Codex and Claude Code, centered on tests, specs, taste, and maintenance.
Key Takeaways
Implement early and rebuild often: cheap code lets you run thought experiments and reconnoiter design space before committing.
Invest in end-to-end behavioral tests, not implementation tests, so frequent rebuilds stay safe and fast.
Keep specs (markdown goals/plans) in sync with code and tests; a frozen pre-work spec misses implementation learnings.
Develop domain taste aggressively: agents amplify existing expertise, and your intuition drives prompt framing, term choice, and debug speed.
Agentic code is “free as in puppies”: generation cost is low but maintenance, support, and security costs are not.
Hacker News Comment Review
Broad consensus that point 10 (maintenance/security costs) is the most important and most ignored, especially as teams chase line-count metrics.
Sharp disagreement on whether “code is cheap” holds at all: critics note current AI pricing is artificially low and engineering complexity is unchanged; the bottleneck shifts to architecture and taste, not keystrokes.
Real-world labor signal from India: junior developer and DevOps hiring has dropped visibly, with campus offers being withdrawn, while senior architectural judgment becomes the scarce resource.
Notable Comments
@mharrison: Argues TDD (tests before code) is the most skipped agentic tip; confirmed it in a live SDK build with no prior SDK existing.
@mxmlnkn: “If you simply can show that you have the patience to maintain a project… you already have an outstanding project” – publish-and-forget is the dominant AI repo pattern.
