Saving Tokens with Script-Based Claude Code Skills

Noticed my /git-sync skill was eating through tokens. Every time I ran it, Claude would read the whole SKILL.md, think about each step, call 5-7 tools… felt wasteful.

So I switched to scripts.

The problem

My original SKILL.md was like 300 lines explaining every step:

# Git Sync Skill

## Step 1: Check Git Status
Run `git status --porcelain` to see changes...

## Step 2: Display Changes
Show the user what will be committed...

## Step 3: ...

Claude would read all of this, interpret it, then execute each step with separate tool calls. Worked fine, but burned through tokens.

What I changed

Created a bash script that does everything:

#!/bin/bash
# git-sync.sh

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

# Show changes
echo "Changes:"
echo "$CHANGES"

# Auto-generate commit message based on changed files
if echo "$CHANGES" | grep -q "skills/"; then
  COMMIT_MSG="feat: Update skills"
elif echo "$CHANGES" | grep -q "docs/"; then
  COMMIT_MSG="docs: Update documentation"
else
  COMMIT_MSG="chore: Update files"
fi

# Create branch if on main
BRANCH=$(git rev-parse --abbrev-ref HEAD)
if [[ "$BRANCH" == "main" ]]; then
  NEW_BRANCH=$(echo "$COMMIT_MSG" | cut -d: -f1)/$(echo "$COMMIT_MSG" | cut -d: -f2 | tr ' ' '-')
  git checkout -b "$NEW_BRANCH"
fi

# Commit and push
git add .
git commit -m "$COMMIT_MSG"
git push -u origin "$(git rev-parse --abbrev-ref HEAD)"

Then simplified SKILL.md to basically just say “run the script”:

# Git Sync Skill

Run this script:
~/.claude/skills/git-sync/git-sync.sh "your commit message"

The difference

Before	After
~300 line SKILL.md	~60 line SKILL.md
5-7 tool calls	1 tool call
~3000 tokens	~500 tokens

That’s like 80% savings per run.

Gotchas

Had a dumb bug where I hardcoded the date in my blog post skill. Wrote 2025-01-16 instead of 2026-01-16.

Fixed it by adding a step to always fetch the system date first:

TODAY=$(date +%Y-%m-%d)

Now the skill instructions explicitly say to get the date from the system, not type it manually.

When to use this approach

Script-based skills work great when:

• The workflow is predictable
• You don’t need Claude to make decisions mid-process
• You run it frequently (token savings add up)

Keep prompt-based skills when:

• You need Claude to analyze/interpret something
• The workflow varies based on context
• It’s a one-off thing

Files structure

~/.claude/skills/git-sync/
├── SKILL.md       # Just says "run the script"
└── git-sync.sh    # All the actual logic

Pretty happy with this. Same functionality, way fewer tokens.