Using AI to Write Better Code More Slowly: A Bug-Finding Workflow
Developers tired of AI slop cannons might appreciate Nolan Lawson's alternative: using LLMs to write better code more slowly. The key insight: throw multiple agents at a PR review to find bugs ranked by severity, then methodically fix them.
How It Works
Lawson describes a Claude skill adapted from another article's core insight: the more different models you throw at a PR review, the fewer hallucinations or bogus bugs you get. The skill runs three agents — Claude sub-agent, Codex, and Cursor Bugbot — to find bugs in a PR, ranked by critical/high/medium/low. After they finish, you review their findings, rule out false positives, and write a final report.
Define "bug" in your own terms: Lawson's includes KISS/DRY principles, accessible HTML/JSX, proper SQL indexes, etc. Claims false positive rate near zero, and the skill always finds tons of bugs — from critical security issues to misleading comments.
Typical Workflow
- Have an agent fix all criticals and highs (with your guidance on proper solution), then repeat until none remain.
- Skip highs/mediums where the fix effort (e.g., 100 lines for a narrow edge case) isn't worth it.
- Abandon the PR if it has so many criticals that the whole approach is misguided.
The review process often finds pre-existing bugs, leading to tangential side-quests writing unit tests and fixing subtle flaws. This is the opposite of 10x productivity slop-cannon development, but improves overall codebase health and deepens your understanding of failure modes.
If you're skeptical of AI coding, this won't persuade you. But if you're churning out multi-hundred-line PRs you barely understand, Lawson invites you to slow down: ask an agent how your PR works and how it might fail, have it write Markdown docs with Mermaid charts, or use Matt Pocock's /grill-me skill until you understand the entire PR front-to-back.
Related discussions on Hacker News: HN thread (748 points, 288 comments).
📖 Read the full source: HN AI Agents
👀 See Also
Using Claude to analyze writing patterns for better custom instructions
A Reddit user describes a method for creating more effective custom instructions by having Claude analyze 10 writing samples to identify concrete patterns like punctuation avoidance and analogy sources, rather than relying on subjective tone descriptions.
One-Soup One-Dish: A Japanese Cooking Principle for AI Fatigue
Takuya applies the Japanese 'Ichiju Issai' cooking principle to combat AI fatigue — simplify your tech stack to one primary tool and one side tool, just like a meal of rice, soup, and one dish.
Leveraging Agent Skills for Writing CUDA Kernels with Upskill
Hugging Face introduces a practical approach to upskill models for writing CUDA kernels using the new Upskill tool, improving model efficiency through agent skills.
Replacing OpenClaw's Default Memory with Redis and Qdrant for Production Multi-Agent Systems
A developer replaced OpenClaw's default SQLite memory with Redis for ephemeral state and Qdrant for persistent vector memory to solve scaling issues in multi-agent setups, implementing semantic search, cross-agent sharing, and concurrent writes.