tokenkiller

What This Skill Does

TokenKiller is a universal throttling and context management skill designed for OpenClaw AI agents. Its primary function is to optimize token consumption across complex multi-step workflows such as debugging, large-scale refactoring, and multi-file explorations. By implementing structured complexity assessments and strict output gating, the skill ensures agents deliver high-value results while preventing runaway token costs.

Key features include the Task Complexity Assessment, which categorizes tasks into Simple, Medium, and Complex tiers to automatically enforce tool call limits and output line budgets. TokenKiller mandates a "Goal First, Evidence Later" approach, utilizes diff-based code patching instead of full file reposting, and requires periodic self-checks every three tool calls. This systematic reduction in redundancy and bloat allows agents to operate efficiently within constrained context windows while maintaining a high success rate.

Installation

To install this skill, use the following command in your terminal: clawhub install openclaw/skills/skills/buttonsinger/tokenkiller

Use Cases

TokenKiller is ideal for developers and power users managing long-running agent sessions. It is specifically recommended when:

• Working with large codebases where reading full files would exceed context limits.
• Performing multi-step debugging sessions that involve deep stack traces or long log files.
• Optimizing API costs for frequent, iterative agent interactions.
• Automating complex refactoring tasks that require multiple sequential file modifications.
• Managing long-term agent memory where context length must be strictly preserved.

Example Prompts

1. "I need to refactor the authentication module across these three files, but please keep the token usage minimal by only showing the diffs."
2. "Debug this 800-line service file, but use TokenKiller to ensure we don't dump the entire log into the context."
3. "Optimize the budget for this migration task; it's a complex multi-module change, so please prioritize concise output."

Tips & Limitations

• Progressive Disclosure: Always fetch only the minimum necessary information. If you need more detail, request it incrementally.
• Diff-First Strategy: Never dump whole files. Rely on the agent to provide patches and command-result summaries.
• Warning System: Pay attention to the "TokenKiller" budget warnings. If an agent hits the limit, switch your prompt to be more specific regarding the target module to regain focus.
• Self-Check: You can manually trigger a self-check by asking the agent if it is adhering to the L0-L2 information layer guidelines.