reedom avatar

reedom-flag-foryou

Document Optimization(Compaction) for Claude Code

Install

/plugin install reedom-flag-foryou@reedom-commands-and-skills

Run it inside Claude Code after adding the marketplace with /plugin marketplace add reedom/claude-code-commands.

Contents

Discovered from source

Skills 1

Context Engineering Optimizer

Optimize content for AI consumption by reducing tokens while preserving essential information.

Core Principle

Context is finite. Every token depletes the attention budget. Optimize ruthlessly.

Reference: context-engineering-guide.md

Optimization Workflow

Use TodoWrite

Step 1: Create Draft

Analyze input and create optimized version:

Remove/Reduce:

  • Redundant explanations (AI already knows common knowledge)
  • Verbose transitions ("In this section, we will discuss...")
  • Duplicate information across sections
  • Filler words and phrases
  • Overly detailed examples when one suffices
  • Charts/diagrams/tables if text conveys same info more efficiently

Preserve:

  • Domain-specific knowledge AI lacks
  • Concrete examples that demonstrate behavior
  • Critical constraints and rules
  • Structural organization (headers, sections)
  • Actionable instructions

Format choices:

  • Prefer bullet points over prose
  • Use imperative form
  • Choose text OR graphical representation, not both
  • Keep code examples minimal but complete

Language:

  • English no matter the original language unless otherwise specified through prompt
  • Most token-efficient for current models
  • Preserves technical terminology accuracy

XML + Markdown tables:

<!-- WRONG -->
<data>
| Col | Val |
|-----|-----|
| A   | B   |
</data>

<!-- CORRECT - empty lines required -->
<data>

| Col | Val |
|-----|-----|
| A   | B   |

</data>

Step 2: Compare and Refine

Compare draft against original:

Check for information loss:

  • Critical details missing?
  • Context for understanding removed?
  • Edge cases dropped that matter?

Check for over-optimization:

  • Can sections be further condensed?
  • Any remaining redundancy?
  • Self-evident comments still present?

Restore if:

  • Meaning becomes ambiguous
  • Important nuance lost
  • Actionability reduced

Step 3: Output

For materials (files to overwrite):

  • Overwrite original with optimized version

For transient content:

  • Output the optimized version as final result

Quality Criteria

Optimized content should:

  • Reduce token count by 30-70% typically
  • Preserve all actionable information
  • Maintain clear structure
  • Be self-contained (no dangling references)
  • Target the Goldilocks zone: specific enough to guide, flexible enough to adapt

Agents 1

Context Engineering for AI Agents

Source: anthropic.com/engineering/effective-context-engineering-for-ai-agents

Core Principle

Context engineering > prompt engineering. Manage the complete token configuration for optimal model behavior, not just instructions.

Why Context Matters

Context rot: More tokens = reduced recall precision (gradient, not cliff) Attention dilution: n^2 pairwise relationships stretch thin as context grows Training bias: Models trained on shorter sequences perform better there

Effective Context Anatomy

System Prompts

Target the Goldilocks zone:

  • Specific enough to guide behavior
  • Flexible enough to avoid brittle logic

Avoid:

  • Hardcoded if-else logic (breaks under variation)
  • Vague high-level guidance (assumes shared context)

Structure with XML tags or Markdown headers: <instructions>, ## Tool guidance

Tools

Well-designed tools are:

  • Self-contained
  • Robust to error
  • Clear about intended use

Failure pattern: Bloated toolsets with overlapping functionality. If humans can't identify which tool to use, neither can agents.

Examples (Few-Shot)

Don't stuff edge cases. Curate diverse, canonical examples that portray expected behavior. Examples function as powerful visual references for LLMs.

Context Retrieval

Just-in-Time (Preferred)

Maintain lightweight identifiers (paths, URLs, queries), retrieve dynamically via tools.

Benefits:

  • Progressive disclosure through exploration
  • File sizes suggest complexity
  • Naming conventions indicate purpose
  • Timestamps signal relevance

Hybrid Strategy

Combine pre-loaded essentials (CLAUDE.md) with autonomous exploration (glob, grep).

Long-Horizon Techniques

Compaction

Summarize conversation history approaching limits. Preserve:

  • Architectural decisions
  • Unresolved bugs
  • Implementation details

Discard:

  • Redundant tool outputs
  • Verbose messages

Approach: Maximize recall first, then optimize precision.

Structured Note-Taking

Write external persistent notes, pull back as needed. Enables multi-hour coherence without memory prompting.

Sub-Agent Architecture

Specialized sub-agents handle focused tasks with clean context. Return condensed summaries (1,000-2,000 tokens) not exhaustive output. Achieves clear separation of concerns.

Key Rules

  1. Context is finite - Every token depletes attention budget. Curate ruthlessly.
  2. Start minimal - Test with minimal prompts on best model, add clarity based on failures.
  3. Efficient tools - Return token-efficient info, encourage efficient behaviors.
  4. Match technique to task:
    • Compaction: extensive back-and-forth
    • Note-taking: iterative development
    • Multi-agent: parallel exploration
  5. Embrace autonomy - Let intelligent models act intelligently with less curation.

Sponsors

Details

Marketplace
reedom/claude-code-commands
Source
flag-foryou
Author
HANAI Tohru
Category
Community

More from reedom/claude-code-commands

All 6