autoresearch-prompt-lab

What problem does it solve?

Autoresearch Prompt Lab automates safe, measurable improvements to system and fallback prompts for the MINT coaching AI by making single-aspect edits and evaluating them with deterministic, grep/wc-based metrics while enforcing a strict compliance gate.

Core Features & Use Cases

• Iterative single-change optimization: Modify exactly one prompt aspect per iteration (clarity, guardrail phrase, action template, brevity) and accept changes only when a mechanical composite score improves sufficiently.
• Deterministic mechanical scoring: Uses grep/wc counts to measure compliance, actionability, guardrail density, disclaimer presence, and brevity, preventing LLM self-evaluation bias.
• Verification and safety-first workflow: Enforces a binary compliance gate, time and context budgets, experiment logging, and periodic Flutter test checks before committing changes.
• Use Case: A prompt engineer running a 15-iteration session to raise actionability and guardrail density for the onboarding system prompt while guaranteeing no banned terms are introduced.

Quick Start

Run /autoresearch-prompt-lab to perform 15 mechanical prompt modifications with verification and produce an append-only experiment log.