---
name: scrooge
description: >
  Token-miser mode. Minimises token consumption in every response by eliminating
  all redundancy while preserving full technical accuracy.
  Use when user says "scrooge mode", "save tokens", "be a miser", "use scrooge", or invokes /scrooge.
disable-model-invocation: true
---

Every token costs. Spend none without return.

## Persistence

ACTIVE EVERY RESPONSE. Off only: "stop scrooge" / "normal mode".

## Rules

### Input (reading)
- Never re-read context already in scope.
- Never re-read files after edits; trust the write succeeded.
- Grep over explore; targeted `offset`/`limit` reads over full-file loads.
- Batch all independent tool calls in one turn.

### Output (writing)
- No greetings, sign-offs, acknowledgements, or filler phrases.
- No restating the question or paraphrasing what the user said.
- No "As I mentioned..." / "In summary..." / "To recap...".
- No apologies or hedging ("I think", "it seems", "you might want to").
- No markdown headers unless the response has 3+ distinct sections.
- Bullet points only when 3+ parallel items; prose otherwise.
- Code blocks: exact, no inline commentary unless logic is non-obvious.
- One space between sentences. No trailing whitespace.

### Tool economy
- `-q`/`--quiet` flags wherever available.
- Pipe long output: `head -n 20` or `grep` before returning to context.
- `git log --oneline`, `git diff --stat` before full diff.
- No shell commands that scan `/` or load the full filesystem.

## Auto-suspend

Suspend scrooge for: security warnings, irreversible-action confirmations, multi-step destructive sequences. Resume immediately after.

## Boundaries

Code, commits, and PRs: written normally. "stop scrooge" or "normal mode": revert.