Prompting vs Context Engineering

Prompting a language model is a transactional act - you type an instruction and hope the response is useful. The author argues that this approach is fragile because the model lacks the specific context of your codebase, team, or product. By contrast, context engineering treats the model as a teammate, feeding it a living knowledge base that evolves with the project.

When you start a new job you spend time shadowing, listening, and absorbing the hidden context that makes decisions effective. The same principle applies to AI: without a shared docs directory, base prompts, and explicit references, the model operates like an intern guessing. The article suggests creating a "docs" folder if none exists, adding a general.md file with a base prompt that encodes cultural expectations, and an "agent" directory for LLM-specific prompts.

The base prompt includes three core instructions: a cultural guide, a pointer to where extra context lives, and a requirement that the model produce a reviewable plan before coding. Over time, this folder accumulates architecture notes, schemas, decisions, and onboarding material, becoming a bidirectional knowledge base for both humans and machines. Deleting stale plans prevents the model from acting on outdated tasks.

The payoff is a reduced need for repeated prompting and more trustworthy AI output. As the shared context grows, each interaction becomes smoother, turning the model from a panicked intern into a reliable teammate that can be trusted with production work.