LLM

• AI Coding Agents Need a Source of Truth

  Bigger prompts do not fix drifting agent plans. A good workflow starts with a small brief, human review, task sizing, and checks against real artifacts.

• How to Cook Claude Code

  Claude Opus 4.7 scores 87.6% on SWE-Bench Verified. Your daily experience probably does not. The gap is not only the model. It is the setup around it.

• It's Not the Model. It's the System Prompt You Never Wrote.

  Most complaints about coding agents are complaints about empty context. CLAUDE.md, distilled project docs, and a few slash commands change the same model from autocomplete into something useful.

• I Asked ChatGPT If Building a Language Was Hard. Now I Have a Compiler.

  Four months ago I asked ChatGPT a dumb question while walking my dog. It told me building a programming language wasn’t that hard. Now I have a compiler, a type system, generics, structured concurrency, and a garbage collector.

• AI Coding Agents Need Memory

  Stateless agents repeat mistakes. lgtm and snap use a findings file, shared codebase context, and recovery loops so later tasks can use what earlier tasks already learned.

• Scaling LLM Codegen Past the One-Shot Demo

  One prompt works for small hacks. Larger projects need product context, technical constraints, module boundaries, sprint-sized slices, and tight test loops.