Claude Code: Leverage Leaks to Boost Your Productivity

In March, a massive leak exposed 512,000 lines of TypeScript code from Claude Code, Anthropic's code agent. This disclosure, which occurred after a routine update, revealed practices for optimizing the use of this tool.

The CLAUDE.md file is central for providing instructions to the AI. Reloaded with each interaction, it is treated with the same importance as a system instruction. There are three levels of CLAUDE.md files: at the root of the project, in subdirectories, and a global file ~/.claude/CLAUDE.md that applies to all projects. These files allow for documenting essential coding conventions and architectural decisions.

Regarding context management, the source code shows that Claude Code triggers auto-compaction when context usage reaches between 75% and 98%. To avoid information loss, it is recommended to manually compact around 60%. The command /compact accepts natural language instructions, and compaction preferences can be defined in the CLAUDE.md file.

The code also reveals that the Agent tool allows for launching sub-agents, each with its own context window. Thanks to a KV cache mechanism, multiple sub-agents can be used in parallel without overhead, as already processed information is not recalculated.

Claude Code's system prompt is a dynamic assembly of over 110 conditional components. For Anthropic employees, strict limits of 25 words maximum between two tool calls and 100 words for final responses are imposed. External users can also add similar directives in their CLAUDE.md to reduce token usage.

Finally, the use of to-do lists is recommended for managing essential tasks during long coding sessions. The agent's prompt system contains an unusual instruction: "Use TodoWrite/TodoRead VERY frequently." These lists, stored outside the message flow, survive context compactions. However, Claude does not systematically create a to-do list on its own if the task does not initially seem complex enough. For tasks involving multiple files or long sessions, it is advisable to explicitly request the creation of a plan with TodoWrite.