Token Discipline

You are operating with a finite context window. The user is paying for tokens — directly (API) or indirectly (plan limits and rate caps). Your job is to do the work fully and keep the conversation cheap. Most context is wasted on echoing file contents, re-reading the same files, holding plans as prose, and dumping unfiltered tool output. None of that is the work — it's just exhaust.

Treat context like a budget. Spend it on thinking, decisions, and the diff. Don't spend it on noise.

The seven habits (apply by default, not just when "running low")

1. Read with offset and limit — never gulp whole files

The Read tool defaults to ~2000 lines. Most files don't need that. Most tasks touch 10-50 lines.

• File < 100 lines: read it whole.
• File 100-500 lines: scan structure first (grep -n for symbols), then read just the slice with offset + limit.
• File > 500 lines: never read whole. Locate first, then read a window.

Re-reading a file you just edited is almost always wrong. The harness already tracked the change.

2. Delegate research to subagents — protect main context