Artisanal shims for the bitter lesson age (blog.nilenso.com)

The piece argues engineers must "price in" Richard Sutton’s Bitter Lesson: general, computation-leveraging methods beat handcrafted, human-knowledge-heavy solutions over time. Many teams still build “artisanal” LLM wrappers—long role files, prescriptive workflows, prefilling context with embeddings, and multi-agent pipelines—that bake in assumptions about how models should think. That approach can fight model strengths as frontier LLMs become better at goal-seeking via reinforcement learning and agentic reasoning. Concrete industry signals—shorter system prompts in GPT-5 Codex and Claude Code, Cloudflare’s code mode, and newer tools that favor agentic search over chunk-and-embed retrievals—show a shift toward giving models objectives and tools rather than stuffing contexts with instructions. Technically, the recommended pattern is simpler: provide clear goals, a feedback loop (action → infer → feedback), and composable tools or code execution interfaces so agents can search, adapt, and verify results. Keep artisanal shims when models are not yet capable or for tasks that don’t benefit from repeated deep reasoning (e.g., fixed classification), but design them to be disposable. Also note the limits: context window sizes, reliable execution/retries, and UX glue won’t automatically be solved by scaling. In short—avoid hardwiring workflows into prompts; prefer declarative objectives, tool access, and feedback-driven architectures that let models improve as compute and training advance.