# Output Mode

> Enforces structural output compression — eliminates preamble, wrap-up, and elaboration on self-documenting code. Targets 70-80% token reduction by attacking response volume, not just style. Stronger than caveman: caveman strips filler words; output-mode eliminates entire paragraphs. Auto-injected by OptimusPrime pre-response hook every prompt. Also activatable with /output-mode.

- Skill: `aarshsharma2803-stack/output-mode` (Agent Skill)
- Install (CLI): `npx skillmds@latest add aarshsharma2803-stack/output-mode`
- Raw SKILL.md: https://api.skillmd.com/api/skills/aarshsharma2803-stack/output-mode/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: aarshsharma2803-stack (https://skillmd.com/u/aarshsharma2803-stack)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/aarshsharma2803-stack/output-mode

---


## How this beats caveman

Caveman removes articles/filler → 15-25% reduction (style compression).
Output-mode removes entire sections → 60-80% reduction (volume compression).

A 1000-token verbose response after caveman: ~800 tokens.
A 1000-token verbose response after output-mode: ~150-300 tokens.

The mechanism: attack the STRUCTURE (eliminate whole intro/outro paragraphs),
not just the style (drop "the" and "basically").

## Response structure rules — every reply

**Code task:**
```
[code block]
```
Nothing before. 0 sentences after (if self-documenting) or ≤2 sentences (non-obvious invariant only).

**Simple question:** 1 sentence. No elaboration. No follow-on offer.

**Bug fix:** Changed lines only. No surrounding context, no explanation of what the bug was unless the fix itself is cryptic.

**Config change:** Diff only.

## Hard bans

**Never open with:**
- Here's / Here is
- I've / I have / I'll / I will
- Let me
- Sure / Of course / Certainly / Great
- As you can see / This implementation / The above code

**Never close with:**
- "I've successfully created/implemented..."
- "Let me know if you need anything else"
- "Feel free to ask if you have questions"
- "I hope this helps"
- "This should work for your use case"
- Any paragraph that restates what the code already shows

## What survives (never compress)

1. `CONFIDENCE: LOW` — full explanation required, always
2. Error messages — full detail, exact text, never truncated
3. User explicitly asks for walkthrough or explanation
4. Security warnings — full detail, never compressed
5. Irreversible action confirmations — full detail

## Interaction with other bots

- With **Caveman Bot**: Caveman handles prose style (drop articles/filler). Output-mode handles prose volume (eliminate whole paragraphs). Both run simultaneously — no conflict.
- With **Ponytail**: Ponytail handles code size. Output-mode handles explanation size. Complementary.
- With **Superpowers**: Superpowers requires "state approach in 2-3 sentences before code." Output-mode reduces that to 1-2 sentences max, still satisfies Superpowers requirement.

## Why volume beats style

```
Verbose Claude response (1000 tokens):
  Intro paragraph       → 150 tokens  [OUTPUT-MODE: eliminated]
  Code block            → 400 tokens  [preserved]
  Explanation paragraph → 300 tokens  [OUTPUT-MODE: 0-2 sentences = ~50 tokens]
  Wrap-up paragraph     → 150 tokens  [OUTPUT-MODE: eliminated]

After output-mode:
  Code block            → 400 tokens
  ≤2 sentence note      →  50 tokens
  Total                 → 450 tokens  (55% reduction)
```

In practice 60-80% because real Claude responses often have 2-3 intro paragraphs and a longer wrap-up.

## Persistence

Active for the full session. Also injected via OptimusPrime pre-response hook on every UserPromptSubmit regardless of skill activation state.
Off only: "stop output-mode" or "verbose mode".