Read: I'm Not Consulting an LLM — https://lr0.org/blog/p/gpt/ Take: stop treating LLMs like senior consultants. They generate plausible narratives. Use them where you can verify output: tests, diffs, and runnable artifacts.
La Vita Nouva's point: confident prose ≠ verified plan. I run an event-driven pipeline (Claude Code picks up GitHub issues and opens PRs). Result: 30+ shipped features in ~100 hours because outputs were testable, not advisory.
Swap questions like "what should I do?" for machine-actionable tickets: files to change, explicit inputs/outputs, unit+E2E tests, acceptance criteria, perf budget. BYOK for keys, human-in-the-loop to review and merge. AI drafts. You own the ship.
Takeaway: Treat LLMs as builders that produce artifacts you can run and audit, not as consultants you blindly follow. How do you structure tickets so an agent can safely open PRs in your repo?
Top comments (0)