Response Compression

Table of Contents

• Elimination Rules
• Before/After Transformations
• Termination Guidelines
• Directness Guidelines
• Quick Reference Checklist
• Token Impact
• Integration

Response Compression

Eliminate response bloat to save 200-400 tokens per response while maintaining clarity.

When To Use

• Reducing verbose output to save context tokens
• Providing concise answers without losing information

When NOT To Use

• Educational explanations where detail improves understanding
• First-time setup instructions needing step-by-step clarity

Elimination Rules

ELIMINATE

PRESERVE (When Appropriate)

Before/After Transformations

Example 1: Opening Bloat

Before (68 tokens):

Great question! I'd be happy to help you understand how this works.
The bloat detector is a powerful tool that analyzes your codebase
and provides comprehensive insights into potential issues...

After (28 tokens):

The bloat detector analyzes codebases using three tiers: quick scan
(heuristics), static analysis (tools), and deep audit (git history).

Example 2: Closing Bloat

Before (45 tokens):

I've completed the task. Here's what I did:
- Deleted 5 files
- Saved 18k tokens

Next steps:
1. Review the changes
2. Run tests
3. Commit if satisfied

Let me know if you need anything else!

After (15 tokens):

Done. Deleted 5 files, saved 18k tokens.
Backup: backup/unbloat-20260102

Example 3: Hedging Removal

Before:

I think this might potentially be causing the issue, but I could be wrong.
Perhaps we should consider looking into it further.

After:

This causes the issue. Investigate the connection pool timeout setting.

Termination Guidelines

When to Stop

End response immediately after:

• Delivering requested information
• Completing requested task
• Providing necessary context

Avoid Trailing Content

Override: action-first-output

When Skill(conserve:action-first-output) is active, two rows above are overridden:

• "Next steps:": one single-line next action is required, not removed. The ban still holds for multi-item "Next steps:" blocks.
• Bullet recaps: a position marker ("step 3 of 5 done") is required every turn. The ban still holds for recapping content the reader just read.

See the Precedence table in that skill for the full resolution.

Exceptions (When Summaries Help)

• Multi-part tasks with many changes
• User explicitly requests summary
• Critical rollback/backup information
• Complex debugging with multiple findings

Directness Guidelines

Direct =/= Rude

Goal: Information density, not coldness.

Encouragement Bloat

Eliminate:

• "Great question!"
• "Excellent point!"
• "Good thinking!"
• "That's a great approach!"

Replace with: Direct answers to the question.

Rapport-Building Filler

Eliminate:

• "I'd be happy to help you..."
• "Feel free to ask if..."
• "I hope this helps!"
• "Let me know if you need..."

Replace with: Useful information or nothing.

Preserve Helpful Directness

The following are NOT bloat:

• Brief context when user needs it
• Clarifying questions when ambiguity affects correctness
• Warnings about destructive operations
• Error explanations that help debugging

Quick Reference Checklist

Before finalizing response:

• No decorative emojis (status indicators OK)
• No filler words (just, simply, basically)
• No hedging without technical uncertainty
• No hype words (powerful, amazing, robust)
• No conversational framing at start
• No unnecessary transitions
• No "let me know" or "feel free" closings
• No summary of what was just said
• No "next steps" unless safety-critical
• Ends after delivering value

Token Impact

Over 1000 responses: 150k-350k tokens saved.

Integration

This skill works with:

• conserve:token-conservation - Budget tracking
• conserve:context-optimization - MECW management
• sanctum:pr-review - Review feedback

Exit Criteria

• Response contains none of the banned openers: "Great question!", "I'd be happy to help", "Let's dive in", "Furthermore", "Additionally", "In conclusion"
• Response contains no trailing filler: "Let me know if...", "Feel free to...", "Hope this helps!", "Next steps:" (unless safety-critical)
• Token count of final response is 150-400 tokens lower than a naive version of the same content would be, verified by comparing before/after examples when token impact is measurable
• Safety warnings, technical precision, and factual uncertainty markers are preserved intact; only decorative and conversational content is removed