Think in Code — 700x Context Savings for AI Agents

The LLM writes the script. The script processes the data. You read only the result.

---

🎯 The Pain Point

Your agent just read 50 TypeScript files to count lines. That's 700KB of context gone on something a 3-line script handles in 3KB.

Sound familiar?

• Agent wants to cat package-lock.json (58,000 lines) to find one dependency
• Agent reads entire access.log (5,000 lines) to count unique IPs
• Agent issues 47 Read() calls to understand dependency frequency
• Context window fills up → compaction triggers → agent forgets what it was doing

Every time an agent reads data into context to process it, you pay a 100-700x context tax.

---

✅ The Solution

Think in Code is a mental model shift + pattern catalog that trains your agent to:

1. STOP reading data into context to process it
2. START writing one-shot scripts that process it
3. READ ONLY the result (3KB vs 700KB)

❌ Anti-pattern:
  Read src/auth.ts → Read src/user.ts → Read src/admin.ts → ... (47 more)
  Cost: ~700 KB context

✅ Think-in-code:
  ctx_execute("bash", `find src -name '*.ts' -exec wc -l {} + | sort -rn | head -20`)
  Cost: ~1 KB output
  Savings: 700x

---

📊 The Numbers

Task	Raw Read	Script Output	Savings
Count lines in 50 .ts files	700 KB	1 KB	700x
Find dependency in package-lock.json	58 KB	< 1 KB	58x
Analyze 5000-line access.log	50 KB	2 KB	25x
Build dependency frequency map (20 files)	300 KB	3 KB	100x
Extract config values from 5 files	15 KB	< 1 KB	15x

Average savings: 50-700x.

---

🚀 Quick Start

Install

# Option 1: Clone directly
git clone https://github.com/aptratcn/skill-think-in-code.git ~/.agent-skills/skill-think-in-code

# Option 2: npx install (skills.sh registry)
npx skills@latest add aptratcn/skill-think-in-code

# Option 3: Add to AGENTS.md (recommended)
echo "## Think in Code\nBefore reading data files: Can a script do this? See ~/.agent-skills/skill-think-in-code/SKILL.md" >> AGENTS.md

Use

Works automatically when the agent recognizes patterns:

Agent: "I need to read package-lock.json to find react version..."
→ Skill activates → Writes script → Reads 3-line result

Or explicitly:

"Use think-in-code to find the top 10 most-used dependencies in this project"

---

🔍 Pattern Catalog (Quick Reference)

JSON Mining

# Find dependencies with vulnerabilities
cat package-lock.json | python3 -c "
import json, sys
data = json.load(sys.stdin)
for name, info in data.get('packages', {}).items():
    if 'vulnerability' in str(info):
        print(name, info.get('version', 'unknown'))
"

Log Analysis

# Top 20 IPs by request count
awk '{print $1}' access.log | sort | uniq -c | sort -rn | head -20

# Top 20 404 paths
grep "404" access.log | awk '{print $7}' | sort | uniq -c | sort -rn | head -20

Dependency Frequency

# Most imported modules across project
grep -rn "import\|require" src/ --include="*.ts" | \
  sed 's/.*import.*from //;s/.*require(//' | \
  tr -d "'\";" | sort | uniq -c | sort -rn | head -30

Config Extraction

# Key values from multiple configs (one shot)
echo "=== TypeScript ===" && grep -A2 "compilerOptions" tsconfig.json && \
echo "=== ESLint ===" && grep "extends\|plugins\|rules" .eslintrc.js && \
echo "=== Jest ===" && grep "testMatch\|coverage" jest.config.ts

---

🛡️ When NOT to Use

Situation	Why
Understanding code architecture	Need full context for mental model
Code review	Need to see patterns, not just data
Writing new code	Need to learn existing patterns
Debugging logic flow	Need to understand the code path

Rule: If you're reading to UNDERSTAND, read. If you're reading to PROCESS, script it.

---

📦 What's Included

skill-think-in-code/
├── SKILL.md          # Full skill instructions (6.8 KB)
├── README.md         # This file
├── LICENSE           # MIT
└── references/
    ├── patterns.md   # 20+ script patterns catalog
    └── examples.md   # Real-world usage examples

---

🔗 Related Skills

Skill	What It Does
skill-audit	Security scanner — audit before install
prompt-guard	Runtime prompt injection defense
token-optimizer	Output compression (~75% reduction)
context-doctor	Diagnose context bloat sources

---

💡 Origin

Inspired by context-mode (11,693 stars, +2,332 this week) — the MCP server that sandboxed tool output.

This skill is the pure markdown version: no npm, no MCP, no dependencies. Just copy the mental model and script patterns into your agent's behavior.

Same 700x savings. Zero install complexity.

---

📄 License

MIT — Use freely, improve freely, share freely.

---

Stop reading. Start scripting. Your context window will thank you. 🧠

⭐ If this skill saved you from a context overflow, consider giving it a star!