Inside `JuliusBrussee/caveman`: A Leaner Claude Code Workflow

Описание и применение

Сокращение контекста в запросах Claude Code до «каменного» стиля снижает потребление токенов примерно на 65 %, повышая экономичность без потери качества задач.

Задача

Repetitive context in Claude Code prompts consumes token quota without improving results.

Результат

Achieved approximately 65% token reduction on routine prompts, improving cost efficiency while maintaining task effectiveness.

Рабочий процесс

• Clone the repository: git clone https://github.com/JuliusBrussee/caveman.git /tmp/caveman
• Create skill directory: mkdir -p ~/.claude/skills/caveman
• Copy skill files: cp -R /tmp/caveman/* ~/.claude/skills/caveman/
• Restart Claude Code
• Send a compact request, e.g., caveman: inspect failed deployment, identify root cause, propose smallest safe fix

Доказательства

The reported 65% token reduction should be treated as workload-dependent rather than a universal benchmark.

Источник

Track token usage before and after adoption across representative repositories, not just toy prompts.

Источник