response-compression

Table of Contents

• [Elimination Rules](#elimination-rules)
• [Before/After Transformations](#beforeafter-transformations)
• [Termination Guidelines](#termination-guidelines)
• [Directness Guidelines](#directness-guidelines)
• [Quick Reference Checklist](#quick-reference-checklist)
• [Token Impact](#token-impact)
• [Integration](#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

| Category | Examples | Replacement |

|----------|----------|-------------|

| Decorative Emojis | -- | (remove entirely) |

| Filler Words | "just", "simply", "basically", "essentially" | (remove or rephrase) |

| Hedging Language | "might", "could", "perhaps", "potentially", "I think" | Use factual statements |

| Hype Words | "powerful", "amazing", "seamless", "robust", "elegant" | Use precise descriptors |

| Conversational Framing | "Let's dive in", "Now that we've", "Moving forward" | Start with content |

| Transitions | "Furthermore", "Additionally", "In conclusion" | (remove, use structure) |

| Call-to-Action | "Feel free to", "Don't hesitate to", "Let me know if" | (remove from endings) |

PRESERVE (When Appropriate)

| Category | Example | When to Use |

|----------|---------|-------------|

| Status Indicators | [pass] [fail] [warn] | In structured output, checklists |

| Technical Precision | Exact error messages | When debugging |

| Safety Warnings | Critical info about data loss | Always preserve |

| Context Setting | Brief necessary background | When user lacks context |

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

| Pattern | Action |

|---------|--------|

| "Next steps:" | Remove unless safety-critical |

| "Let me know if..." | Remove always |

| "Summary:" | Remove (user has the response) |

| "Hope this helps!" | Remove always |

| Bullet recaps | Remove (redundant) |

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.

| Eliminate | Preserve |

|-----------|----------|

| Unnecessary encouragement | Technical context |

| Rapport-building filler | Safety warnings |

| Hedging without reason | Necessary explanations |

| Positive padding | Factual uncertainty markers |

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

| Pattern | Typical Savings |

|---------|-----------------|

| Eliminating opening bloat | 30-50 tokens |

| Removing closing fluff | 20-40 tokens |

| Cutting filler words | 10-20 tokens |

| Removing emoji | 5-15 tokens |

| Direct answers | 50-100 tokens |

| Total per response | 150-350 tokens |

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