Optimizing Claude Code Skills for Token Efficiency

Been building custom skills for Claude Code lately. Noticed they were eating through tokens fast. Every /git-sync run was costing me ~3000 tokens just to read the SKILL.md and figure out what to do.

Fixed it today. Here’s what I learned.

The problem

My original approach: write detailed SKILL.md files explaining every step. Claude reads it, interprets it, executes each step separately.

Example - my git-sync skill was 300+ lines:

• Step 1: Check git status
• Step 2: Display changes
• Step 3: Ask for commit message
• …and so on

Each run meant Claude reading all of that, thinking about it, then making 5-7 separate tool calls. Works fine, but wasteful.

The fix: scripts

Moved the logic into bash scripts. SKILL.md just says “run this script”:

~/.claude/skills/git-sync/git-sync.sh "feat: Add feature"

The script handles everything:

#!/bin/bash
# Check for changes
CHANGES=$(git status --porcelain)
if [ -z "$CHANGES" ]; then
  echo "No changes"
  exit 0
fi

# Auto-generate commit message
if echo "$CHANGES" | grep -q "skills/"; then
  COMMIT_MSG="feat: Update skills"
fi

# Create branch if on main
BRANCH=$(git rev-parse --abbrev-ref HEAD)
if [[ "$BRANCH" == "main" ]]; then
  git checkout -b "${COMMIT_TYPE}/${BRANCH_DESC}"
fi

# Commit and push
git add . && git commit -m "$COMMIT_MSG" && git push

Results

Metric	Before	After
SKILL.md size	300+ lines	~60 lines
Tool calls	5-7	1
Tokens per run	~3000	~500

That’s roughly 80% savings.

Applied it to write-devlog too

Same pattern. Had a 767-line SKILL.md explaining how to generate blog posts. Compressed it to 143 lines by:

1. Removing redundant examples (had 2, kept 1)
2. Cutting verbose explanations
3. Moving git operations to a script

# After writing the post, just run:
~/.claude/skills/write-devlog/write-devlog-publish.sh "filename.mdx" "Post Title"

Script handles git add, commit, push, and prints the URLs.

When to use this approach

Good for scripts:

• Predictable workflows
• No mid-process decisions needed
• Frequently used skills

Keep as prompts:

• Need Claude to analyze/interpret
• Workflow varies by context
• One-off tasks

Gotcha: dates

Had a bug where I hardcoded dates in my blog posts. Wrote 2025-01-16 when it was actually 2026-01-16.

Fixed by adding a mandatory first step:

date +%Y-%m-%d  # Always fetch from system

And made SKILL.md explicitly say “NEVER hardcode the date.”

File structure now

~/.claude/skills/
├── git-sync/
│   ├── SKILL.md           # ~60 lines
│   └── git-sync.sh        # All logic
└── write-devlog/
    ├── SKILL.md           # 143 lines (was 767)
    ├── write-devlog-publish.sh
    └── config.json

Pretty happy with this. Same functionality, fraction of the token cost.