prompt slimmer

Prompt bloat is a recurring tax on every LLM call. This skill reduces token counts by 30-60% while ensuring behavior remains identical through rigorous verification. It eliminates the risk of silent regressions by using a structured loop of measurement, incremental cuts, and output-equivalence testing. Whether you are managing a massive system prompt, stuffed RAG contexts, or redundant few-shot examples, this tool provides the discipline to trim costs without degrading quality.

What it does

• Token reduction identifies and removes redundant rules, hedging prose, and format specs the model already follows.
• Behavior verification proves output equivalence using a jury of LLM judges or programmatic eval suites.
• Regression gating separates real behavior shifts from LLM randomness using a flaky-test triage process.
• Cost reporting calculates exact token savings and projected monthly ROI based on call volume.

How it works

1. Measure the baseline token count and capture current behavior across a representative input set.
2. Cut in small, 5% increments, focusing on low-risk areas like hedging or restated rules.
3. Verify using the output-equivalence council for open-ended text or the triage gate for checkable outputs.
4. Iterate by keeping verified shrinks and reverting any edit that causes a behavior drift.

Frameworks & tools

This skill includes standalone Python 3.9+ scripts for CLI-based equivalence checking and test triaging. It is provider-agnostic and works with any testing framework that produces parseable failure lines, including pytest, vitest, and jest.

Why this beats prompting it yourself

Manual prompt trimming often leads to "disguised guard" failures where a seemingly minor deletion breaks a specific edge case. This skill replaces guesswork with a mathematical loop that requires empirical proof before any cut is finalized.

Use cases

• Shrinking a massive system prompt to fit within a smaller model's context window.
• Reducing the monthly API bill for high-volume RAG or agentic workflows.
• Adding a prompt-regression gate to a CI/CD pipeline to prevent cost bloat.
• Auditing few-shot examples to find the minimum set required for accuracy.

Known limitations

This skill is strictly for shrinking prompts without changing behavior. It does not support adding new features or fixing bugs during the slimming process.