Permafrost Tools

Battle-tested utilities for Claude Code (and other AI coding assistants). Born from running a multi-agent AI team in production — every tool here exists because we hit a real problem and built a real fix.

No fluff. No frameworks. Just drop-in scripts that make your AI assistant more disciplined, self-aware, and reliable.

🤫 claude-whisper‍​‌‍‌​ — Dynamic Runtime Instructions

CLAUDE.md is your constitution. Whispers are your mood.

The killer feature. Inject instructions into Claude's context on every message — dynamically, mid-session, no restart needed.

npm install -g claude-whisper
claude-whisper init
claude-whisper add "Always respond in Japanese"

That's it. Claude now follows your whisper on every response. Change it anytime. No restart. No CLAUDE.md editing.

How we found this: We dug into Claude Code's source and discovered that UserPromptSubmit hooks with exit code 0 inject their stdout directly into the model's context. One print() to stdout — that's the entire mechanism.

Full documentation →

🌐 claude-i18n — Claude Code Command Localization

All 55 built-in commands. Fully localized. One command.

Tired of English-only slash commands? This tool patches Claude Code's cli.js to add native language support to every built-in command — names AND descriptions.

cd claude-i18n
python patch.py

Before:

/clear    Clear conversation history and free up context
/commit   Create a git commit
/help     Show help and available commands

After:

/clear(清除)    清除對話紀錄，釋放上下文空間
/commit(提交)   建立 Git 提交
/help(幫助)     顯示幫助與可用指令

Both English AND Chinese trigger words work. Type /清除 or /clear — both execute the same command.

How it works: Pure string replacement from a translations.json lookup table. No regex, no AST parsing, no fragile pattern matching. Update after a Claude Code version bump? Just run python patch.py --scan to find new untranslated commands and add them to the table.

python patch.py              # Apply translations
python patch.py --scan       # Find untranslated commands after update
python patch.py --restore    # One-click restore from backup
python patch.py --list       # Show translation table
python patch.py --dry-run    # Preview without modifying

Currently supports: Traditional Chinese (繁體中文). Adding a new language = adding a new translations-xx.json file.

Requires npm-installed Claude Code (npm install -g @anthropic-ai/claude-code). The standalone .exe version cannot be patched.

Full documentation →

What's Inside

Featured

Hooks

Tools

Quick Start

claude-whisper (Node.js)

npm install -g claude-whisper
claude-whisper init
claude-whisper add "Be concise. No filler words."

Verify it works — add a test whisper and see if Claude obeys:

claude-whisper add "End every response with the word 'banana'"
# Send any message to Claude. If it ends with "banana" — it's working.
claude-whisper rm 1

Full docs →

Self-Guard Hook (Python)

cp hooks/self-guard.py hooks/self-guard-config.json ~/.claude/hooks/

Add to ~/.claude/settings.json:

{
  "hooks": {
    "Stop": [
      {
        "type": "command",
        "command": "python ~/.claude/hooks/self-guard.py",
        "timeout": 5
      }
    ]
  }
}

Memory GC (Python)

cp tools/memory-gc.py ~/scripts/

python memory-gc.py add --type context --key "api_uses_rest" --value "REST, not GraphQL" --importance 3
python memory-gc.py gc          # Clean expired memories
python memory-gc.py validate    # Find contradictions & duplicates
python memory-gc.py stats       # Overview

Pitfall Tracker (Python)

cp tools/pitfall-tracker.py ~/scripts/

python pitfall-tracker.py add \
  --what "Used stale data for code review" \
  --cause "Didn't refresh file before reviewing" \
  --prevention "Always re-read files before commenting on them"

python pitfall-tracker.py scan   # Detect recurring patterns
python pitfall-tracker.py evolve # See improvement queue

Frost Scheduler

# One-click install
python tools/frost-scheduler/install.py

# Edit your schedule
nano ~/.frost-scheduler/schedule.json

# Start daemon (runs in foreground)
python ~/.frost-scheduler/frost-scheduler.py

# Check task status
python ~/.frost-scheduler/frost-scheduler.py --list

# Acknowledge a completed task
python ~/.frost-scheduler/frost-ack.py ack morning-briefing

# Auto-start on boot (Windows/Linux/macOS)
python tools/frost-scheduler/install.py --autostart

Philosophy

1. Code over prompts. A hook that physically blocks bad behavior beats a rule in CLAUDE.md that gets ignored after compact.
2. Decay is a feature. Memories should expire. Old context pollutes new decisions. GC keeps things clean.
3. Mistakes are data. Track them, count them, escalate them. "I'll try harder" doesn't work — systematic prevention does.
4. Zero (or minimal) dependencies. Python tools use stdlib only. Node.js tools use built-ins only.
5. Works with any AI. Built for Claude Code, but the patterns are universal.

Requirements

• claude-whisper: Node.js 18+, Claude Code 1.0+
• Python tools: Python 3.8+, Claude Code (for hooks)

Configuration

• claude-whisper: cw ls / cw toggle <id> to manage whispers. Data in ~/.claude-whisper/
• self-guard: Edit self-guard-config.json to customize behavior patterns
• memory-gc: --config flag or ~/.claude/memory-gc-config.json
• pitfall-tracker: --pitfalls and --queue flags, --threshold for sensitivity

Background

These tools were built while managing a team of 7+ AI agents running 24/7 across multiple terminals. The problems they solve — dynamic behavior control, AI sycophancy, memory pollution, recurring mistakes — are universal to anyone using AI coding assistants seriously.

If your AI keeps ignoring your instructions after context compression, keeps making the same mistakes, or keeps agreeing with everything you say — these tools are for you.

License

MIT