Context

To ensure the coding agents perform to the highest possible standard, their context should focus on the current problem. The following strategies can help you keep the context concise.

Manage the context proactively

  • Ensure that the current token usage is displayed at all times.

  • Clear the context when you switch to a different task. You usually also have the option to save the context beforehand so that you can refer back to it later. If necessary, you can provide instructions on what should be retained in the summary.

Reduce the overhead of the MCP server

  • MCP tool definitions are usually deferred, so that only the tool names are available in the context until the coding agent uses a specific tool.

  • Where possible, use CLI tools – they are still more context-efficient than MCP servers, as they do not add tool-specific entries. Coding agents can execute CLI commands directly.

  • Disable unused MCP servers.

Install language-specific plugins

These plugins enable coding agents to navigate symbols precisely, rather than relying on text-based searches, thereby reducing unnecessary file reads when browsing through unfamiliar code. Installed language servers also automatically report type errors following changes, allowing coding agents to identify errors without having to run a compiler. In Claude Code, this is pyright-lsp for the Pyright Language Server.

Shift the work to hooks and skills

  • Custom hooks can pre-process data before passing it on to coding agents. Instead of having the coding agent search log files for the ERROR log level and return only the relevant lines, thereby reducing the context from tens of thousands of tokens to hundreds.

  • Agent skills provide domain knowledge so that the coding agent does not have to research it itself. A codebase-overview skill could, for example, describe your project’s architecture, key directories and naming conventions. When the coding agent calls the skill, it receives this context immediately, rather than having to spend tokens reading through multiple files to understand the structure.