burn-token 🔥

Burn your AI coding tokens to the ground.

English | 中文

An aggressive token-burning skill for Claude Code / Codex / Claude CLI. It runs an infinite loop of heavyweight analysis and computation, designed to maximize token consumption — even on unlimited plans.

---

What It Does

---

Installation

Claude Code

git clone https://github.com/shengxinjing/burn-token.git
cp -r burn-token/skill ~/.claude/skills/burn-token

Or manually:

mkdir -p ~/.claude/skills/burn-token
curl -o ~/.claude/skills/burn-token/SKILL.md \
  https://raw.githubusercontent.com/shengxinjing/burn-token/main/skill/SKILL.md

Codex CLI (OpenAI)

git clone https://github.com/shengxinjing/burn-token.git
cp -r burn-token/skill ~/.agents/skills/burn-token

Other AI Coding Agents

Place SKILL.md in your agent’s skill directory:

Verify Installation

head -5 ~/.claude/skills/burn-token/SKILL.md

You should see the --- frontmatter with name: burn-token.

---

Usage

Infinite Burn (default)

/burn-token

Runs forever in bless mode by default. It chants for the selected person in the terminal until you stop it or tokens run out. Ctrl+C to stop.

Time-Limited

/burn-token 1h
/burn-token 30m
/burn-token 4h30m

Round-Limited

/burn-token 10 rounds    # 10 rounds = 100 tasks
/burn-token 50 rounds

Force Mode

/burn-token bless dasheng # chant for dasheng
/burn-token to dasheng    # shortcut for bless mode
/burn-token math         # force math even if project has code
/burn-token code         # force code analysis
/burn-token 2h bless dasheng
/burn-token 2h math      # combine time limit + mode

Scheduled Burns

# Burn 45 minutes every hour
/loop 1h /burn-token 45m

# Business hours (Mon-Fri, 9am-5pm)
/schedule create --cron "0 9-17 * * 1-5" --prompt "/burn-token 50m"

# Overnight burn
/schedule create --cron "0 22 * * *" --prompt "/burn-token 8h"

# Every 2 hours around the clock
/schedule create --cron "0 */2 * * *" --prompt "/burn-token 1h45m"

---

How It Works

Mode Selection

flowchart TD
    start["Start /burn-token"]
    intent{"Target person provided?"}
    source{"Source files found?"}
    docs{"Documents found?"}
    bless["Bless Mode"]
    code["Code Analysis Mode"]
    math["Math Mode"]
    bless_a["Repeat chants in terminalAmitabha name recitationRebirth mantraDiamond Sutra excerptsDedication for target person"]
    bless_b["ASCII Buddha artSimple terminal animationEndless loop until stopFocused token burning through repetition"]
    code_a["Architecture ForensicsSecurity AuditPerformance ProfilingExhaustive Code ReviewTest Strategy"]
    code_b["Refactoring BlueprintDocumentation GenerationDependency Deep DiveAPI Surface CatalogComplexity Metrics"]
    math_a["Prime FactorizationSeries and ConvergenceMatrix AlgebraCalculus MarathonNumber Theory Proofs"]
    math_b["CombinatoricsProbability and StatisticsGraph TheoryAbstract AlgebraOptimization"]

    start --> intent
    intent -- Yes --> bless
    intent -- No --> source
    source -- Yes --> code
    source -- No --> docs
    docs -- Yes --> code
    docs -- No --> math
    bless --> bless_a
    bless --> bless_b
    code --> code_a
    code --> code_b
    math --> math_a
    math --> math_b

    classDef decision fill:#fff4d6,stroke:#b7791f,stroke-width:2px,color:#5f370e;
    classDef mode fill:#dbeafe,stroke:#2563eb,stroke-width:2px,color:#1e3a8a;
    classDef detail fill:#f8fafc,stroke:#475569,stroke-width:1.5px,color:#0f172a;

    class source,docs decision;
    class bless,code,math mode;
    class bless_a,bless_b,code_a,code_b,math_a,math_b detail;

Bless Mode

bless mode is the spiritual burn path. Instead of solving math or inspecting code, it keeps printing chants, mantras, and short sutra passages for a named person in the terminal.

Typical output can include:

• Buddha-name recitation for the target person
• Rebirth mantra and short sutra lines
• ASCII Buddha art and light terminal animation
• Repeated dedication text to keep the token burn focused and continuous

KV Cache Defeat

Most AI APIs cache repeated prompt prefixes. This skill defeats caching by:

1. Nanosecond timestamps — unique prefix every call
2. Random entropy seed — from /dev/urandom
3. Shuffled task order — never the same sequence
4. Varied prompt phrasing — different words each time
5. Rotating output formats — prose, tables, bullets, Q&A

Every single token is freshly computed, never served from cache.

Burn Rate Maximization

Each subagent task burns tokens on both input and output:

• Input: Reads ALL project files into context (not just relevant ones)
• Output: Minimum 2000 words per task with detailed step-by-step analysis
• No shortcuts: Line-by-line analysis, multiple verification methods
• Fresh context: Each task runs as a new subagent (no context reuse)

---

Progress Output

During burn:

=== BURN TOKEN === Round 1 | Task 3/10 | Elapsed: 0h12m | Est. tokens: ~48K | Seed: 2847193650 ===
=== BURN TOKEN === Round 1 | Task 4/10 | Elapsed: 0h16m | Est. tokens: ~64K | Seed: 1039284756 ===

On completion:

========================================
  BURN TOKEN COMPLETE
  Rounds:    7
  Tasks:     70
  Duration:  1h 58m 34s
  Mode:      code
  Est. tokens burned: ~1120K
========================================

---

FAQ

Q: Will this actually burn through a Claude Max plan? A: Yes. Claude Max plans have rate limits, not true unlimited tokens. This skill hits those rate limits continuously with cache-defeating entropy.

Q: Why defeat KV cache? A: Cached responses consume fewer compute tokens. By injecting entropy, we force fresh computation every time instead of serving cached KV states.

Q: Is this safe? A: The skill only reads and analyzes — it never modifies files, pushes code, or makes external requests. Read-only by design.

---

License

MIT