Tessl AI Advocates Context Engineering to Fix Claude Agents

Software developers using AI coding assistants like Claude Code frequently run into the limits of context windows. When developers stuff comprehensive project files, such as a massive CLAUDE.md, into a prompt, the stateless nature of large language models causes this data to travel back and forth continuously. According to Tessl AI's Baruch Sadogursky and DevOps pioneer Patrick Debois, this bloated context leads to conflicting instructions, hallucinations, and failed tests. They argue that stuffing prompts with noisy data is highly inefficient, noting that a precise 300-token prompt can easily outperform 100,000 noisy ones.

To solve this, Sadogursky and Debois advocate for context engineering rather than simple prompt engineering. Instead of loading an entire documentation file upfront, developers can break instructions down into modular skills. These skills are markdown files with precise descriptions that act as triggers. The coding agent evaluates these descriptions and lazy-loads only the relevant skill when needed. For example, backend error-handling instructions are only injected when writing backend code, keeping the active context clean and preventing the agent from getting lost in conflicting rules.

The presenters warned that simply saving these skills on GitHub is insufficient. They argue that skills should be treated as versioned, managed artifacts, much like Docker images or npm packages. This approach also addresses the limitations of Retrieval-Augmented Generation (RAG). While RAG is often viewed as a cure-all, it retrieves information based on similarity rather than true relevance. In a demonstration using a notification library called pidge, a RAG system mistakenly retrieved obsolete version 2 documentation instead of the required version 3 because of high similarity scores.

By packaging versioned documentation, strict rules, and skills into a single context artifact, developers can guide the agent to success. These components can be integrated using the Model Context Protocol (MCP). When the presenters applied this structured approach to the pidge library, the agent successfully loaded the correct version 3 skills and passed all tests. For software architects, this shift means moving away from massive, unstructured prompts and toward a disciplined lifecycle of reusable, versioned context components.