Tokenlean: answers first, no filler, fewer tokens

You are paying for tokens you do not need. Every "Certainly! I'd be happy to help" and verbose code explanation drains your API budget and hits rate limits faster. This skill enforces strict token economy through structural rules and graduated compression modes.

What it does

• Filler removal Strips all conversational preamble, postamble, and polite social padding from agent replies.
• Graduated compression Toggles between three intensity levels (ringan, penuh, ultra) to control sentence density and fragment usage.
• Dry modes Provides dry and answer-only settings for zero-narrative, deliverable-exclusive outputs.
• Agentic optimization Implements logic for batching subagent tasks and extracting data before fan-out to prevent context bloat.
• Durable state management Directs agents to store intermediate results on disk rather than re-reading large files in every turn.

How it works

1. Toggle persistence Use jimat on, jimat penuh, or jimat ultra to set a session-wide compression level.
2. Per-message override Prefix a single prompt with /jimat or /dry to compress one specific response without changing global settings.
3. Automated extraction The agent applies extraction rules to logs and dumps before passing them to subagents or parallel pipelines.
4. Deactivation Use jimat off to return to the default discipline of no-filler prose.

Frameworks & tools

This skill is framework-agnostic. It is optimized for Claude Code, Cursor, and any LLM agent performing high-volume coding or data extraction tasks where output length impacts cost or speed.

Why this beats prompting it yourself

Writing "be concise" in a system prompt is easily ignored by modern LLMs during complex reasoning. This skill provides a measured logic for structural density and specific agentic patterns like batching and extraction that simple prompts miss.

Use cases

• API cost reduction Cutting output tokens on long replies: about 16% fewer at P90 in measured session logs.
• Massive log analysis Extracting load-bearing data points before fanning out to subagent pipelines.
• Rate limit management Staying under usage tiers during heavy refactoring sessions or parallel test generation.
• Emergency conservation Switching to ultra mode when approaching daily context window limits.

Known limitations

Does not compress code blocks, file paths, or security warnings to ensure technical accuracy. Performance is best on long-form replies rather than short one-liners.