Version: 1.1.0 | Author: Shadows Company | License: MIT

WHEN TO TRIGGER

• Context window approaching capacity (>70% usage)
• Agent responses becoming slower or less coherent
• User says "optimize context", "compact", "reduce tokens"
• Working with very large codebases
• Multi-file operations causing context bloat

WHEN NOT TO TRIGGER

• Short conversations with plenty of context remaining
• Simple single-file operations

PREREQUISITES

No binaries required. This is a pure reasoning skill about optimizing context window usage. It provides strategies and patterns — it does not execute commands or access external systems.

PRINCIPLES

1. Reference Over Inline

Instead of reading entire files into context, reference them:

• "The auth module at src/auth/index.ts handles..." instead of pasting 500 lines
• Read only the specific functions/sections needed
• Use line ranges: Read file.py lines 45-80 instead of the whole file

2. DRY Prompts — Zero Duplication

• Never repeat information already in system context
• Don't re-describe tools you already know about
• Don't re-state project conventions that are in CLAUDE.md/SOUL.md
• If a fact was established earlier, reference it, don't restate it

3. Lazy-Load Strategy

• Load detailed context only when needed for the current task
• Use subagents/sub-tasks for exploration (protects main context)
• Delegate research to agents, keep main context for execution

4. Smart File Reading

WRONG: Read the entire 2000-line file
RIGHT: Read lines 150-200 where the function is defined
WRONG: Read all 15 config files
RIGHT: Read only the config relevant to current task
WRONG: Grep the entire codebase for "import"
RIGHT: Grep specific directories for specific patterns

5. Output Compression

When reporting results:

• Lead with the answer, not the reasoning
• Skip filler words and unnecessary transitions
• Use tables for comparative data
• Use bullet points for lists, not paragraphs

TECHNIQUES

Technique 1 — Context Audit

Assess current context usage:

• Count how many files have been read in this session
• Identify which file contents are still relevant to the current task
• Determine what information can be summarized instead of kept verbatim
• Flag redundant tool results that repeat already-known information

Technique 2 — Compaction

When context is high:

• Summarize completed work (keep outcomes, drop process details)
• Drop file contents that are no longer needed for the active task
• Keep only active task context in working memory
• Preserve critical state: decisions made, errors encountered, current objectives

Technique 3 — Subagent Delegation

For research-heavy tasks:

• Spawn a subagent for codebase exploration
• Subagent returns only findings (not raw file contents)
• Main context stays clean for implementation
• Multiple subagents can run in parallel for independent queries

Technique 4 — Structured Responses

WRONG (100 tokens):
"I've looked at the file and after careful analysis I believe that
the issue is related to the authentication middleware where the
token validation function doesn't properly handle expired tokens."
RIGHT (30 tokens):
"Bug: validateToken() in auth middleware doesn't handle expired
tokens. Fix: add expiry check at line 45."

ANTI-PATTERNS TO AVOID

RULES

• Minimum viable context — load only what's needed NOW
• Summarize, don't accumulate — compress completed work
• Delegate exploration — use subagents for research
• Direct answers — skip preamble, lead with the point
• 3-search maximum — never use more than 3 search tools for one query

SECURITY CONSIDERATIONS

This skill is purely advisory — it provides strategies for token optimization. It does not execute commands, read files, make network calls, modify configuration, or store data. Zero risk profile.

• Commands executed: None
• Data read: None (advisory reasoning only)
• Network access: None
• Persistence: None
• Credentials: None required
• File modification: None

OUTPUT FORMAT

Apply the techniques above inline during agent operation. No separate report is generated — the skill manifests as improved efficiency in the agent's behavior: shorter responses, fewer tool calls, targeted file reads, and minimal context consumption.

Published by Shadows Company — "We work in the shadows to serve the Light."