Use compact output flags for common Bash commands to reduce token consumption

Summary

Claude Code's system prompt already instructs the model to prefer dedicated tools (Read, Grep, Glob) over Bash equivalents. But commands that must go through Bash — git, test runners, build tools — still use verbose default output, consuming significant context tokens unnecessarily.

The model could be instructed to prefer native compact output formats when generating Bash commands. This is zero-overhead, zero-dependency, and preserves all error information.

Proposed flags

| Command | Compact flag | Effect |
|---------|-------------|--------|
| git status | --porcelain | Machine-readable, no prose |
| git log | --oneline | One line per commit |
| git diff | Prepend --stat | File-level summary before full diff |
| pytest | --tb=short -q | Compact tracebacks, quiet pass output |
| cargo test | Pipe to tail -20 | Summary only unless debugging |

Why this works

These are upstream optimizations — asking the tool for less noise rather than post-processing output. Key properties:

• No information loss on errors (compact modes preserve failure details)
• No external dependencies or shell hooks
• No latency overhead
• Implementable as a system prompt change

Context

Analysis of RTK (a CLI proxy for token compression) shows that injecting upstream flags is their highest-leverage strategy — responsible for ~70% of savings. The remaining filtering/truncation techniques involve lossy compression with real tradeoffs. The flag approach avoids those tradeoffs entirely.

---
🤖 Written by Claude Code