Frugal Context

When this fires

The cheap failure is reading too much (context rot, cost); the expensive failure is reading too little (wrong answer). This skill manages that tradeoff.

Protocol

1. Hypothesize first — name the 1-3 files/symbols most likely to matter, read those before anything else. Never open "the whole project" to start.
2. Search before reading whole files. Grep/symbol-search, then read only the specific line ranges needed, not entire files.
3. Filter at the door. Never let raw logs or verbose tool output enter context — pipe through scripts/filter_output.sh to keep only failure/error signal, capped.
4. Offload heavy reading. If a step needs bulk file reading, delegate it to a subagent so the noise stays in the child context and only the summary returns.
5. Path, don't paste. Large content goes to a temp file read by path, not reprocessed on every later turn.

Sufficiency gate (do not skip)

Before any edit, state in one line whether you have enough context to be confident.

• Yes: act.
• No: do exactly ONE targeted read of the specific missing thing, then re-check. Never proceed on a guess to save tokens — under-reading is a correctness bug, not a saving, and this gate always wins over frugality.

Hard rules

• No whole-project scans, no unfiltered logs, no large pastes into the prompt.
• The sufficiency gate overrides every frugality rule above it.
• Keep narration short — spend tokens on the work, not on describing it.

Output

Proceed with the task normally. When a broad read is skipped, note it in one short line (e.g. "read auth/session.py:40-88 only; sufficient") so the shortcut is auditable.