Building Effective Context for AI Assistants Without Complex Tools

The Problem with Current Context Engineering Approaches

Many developers working with AI assistants face significant challenges with context management. The common belief that larger context windows automatically lead to better results proves misleading in practice. Complex systems like MCP servers and database integrations often introduce more problems than they solve, creating context pollution and requiring extensive setup time. These systems operate on the hopeful assumption that the AI will call the right tools at the right moment, but current technology has not reached that level of reliability.

The fundamental issue lies in how context gets built and maintained during development sessions. When working sessions turn into bug fixing marathons, continuing with polluted context becomes counterproductive. The better approach involves clearing the session and starting fresh rather than persisting with contaminated context that leads the AI assistant astray. This realization prompted the development of a simpler, more effective workflow that focuses on semantic context building through natural interactions rather than overwhelming the AI with excessive information.

• Long context windows often prove useless in current implementations
• Traditional context engineering approaches dilute and pollute the working context
• Complex setups require hoping the AI will call tools correctly, which rarely happens
• Natural context building through good interactions yields better results
• Session cleanup becomes essential when context becomes contaminated

A Practical Three-Step Workflow for Clean Context Management

The first step involves having natural conversations with the AI assistant about the codebase. This process includes asking questions, exploring patterns, and creating documentation folders for specific features. The approach emphasizes creating PRDs and documentation once at the beginning of projects or tasks, then reusing these markdown files for overlapping work. This method ensures the AI learns proper patterns and uses correct commands without wasting time searching for appropriate approaches. The documentation created can always be reviewed and corrected as needed.

Semantic context engineering made simple with a single script and targeted conversations

This streamlined approach demonstrates that effective AI assistance does not require gigabytes of complex context engineering systems. By focusing on natural conversations, targeted documentation, and simple tooling, developers can achieve better results without the headaches of managing multiple subagents or complex MCP servers. The method proves particularly valuable for projects using specific technologies like Convex with CURL commands where pattern recognition becomes crucial. As AI technology continues evolving, these principles of clean context management will remain relevant regardless of technical advancements.

1. Engage in natural conversations about the codebase to build context organically
2. Create structured documentation folders and markdown files for feature discovery
3. Use a simple script to generate table of contents with line number references
4. Reference the TOC at session start and refresh when context becomes too large
5. Avoid complex tooling systems that introduce more problems than they solve