What problem does it solve? List-building tools return thousands of companies without knowing whether they match your ideal customer profile, so enriching and emailing unqualified companies wastes budget. This Skill tunes an AI qualification prompt on small samples before you scale, so only true ICP fits move forward. ## Core Features & Use Cases - Iterative tuning loop: Scores batches of 10 companies via Claude Code Task sub-agents, collects your corrections, refines the prompt, and stops after 2 consecutive zero-correction rounds. - No external API keys: All scoring runs inside Claude Code with Task sub-agents, so no Anthropic or OpenAI key management is required during tuning. - Reusable output: Saves the converged prompt to a profile file and records metadata in client-profile.yaml, then applies it at scale with scripts/score-batch.ts. - Use Case: After pulling 5,000 companies with a list-builder, tune the prompt on 30-50 samples, lock it in, and filter the full list to only qualified companies with confidence >= 0.6 before uploading to Smartlead. ## Quick Start Run the icp-prompt-builder skill after a list-building skill to tune an ICP qualification prompt on a sample of 10 companies from my results.
