The token-saving skill enforces concise, low-noise AI responses. It is always active — it applies to every reply without needing to be triggered by a specific phrase.
Without the skill, AI assistants commonly pad responses with filler openers, closing platitudes, and repeated context. This skill removes that noise.
| Behaviour | Without skill | With skill |
|---|---|---|
| Response opener | "Great question! Certainly, I'd be happy to help..." | Directly answers |
| Closing | "Let me know if you have any questions!" | Stops when done |
| Repeated context | Restates what you just said | Skips it |
| Factual answers | Unlimited prose | ≤ 5 lines |
| Action lists | Unlimited bullets | Capped at 4 |
| Full file dumps | Common | Only when you ask |
Every response that writes or changes code ends with exactly this footer — no more, no less:
**What changed:** <one line>
**Why:** <one line>
**How to verify:** <command or test instruction>
This applies to: new functions, patches, inline diffs, config snippets, or any code block representing a change.
It does not apply to: Q&A, reviews, planning, comparisons, or conceptual explanations.
The brevity rules suspend for a single response when you explicitly ask for depth:
"Give me a full explanation." "Deep dive into this." "Don't hold back." "Complete explanation."
The next response is fully unrestricted. Rules resume after that.
If another active skill specifies its own output format (e.g. a review skill with a Blocker / Important / Nit structure), that format takes precedence over this skill's rules.
See Getting Started for the full install guide.