AI Outreach Automation — AI skill for Claude Code

AI Outreach Automation

A [Claude Code](https://claude.com/claude-code) **skill** that turns a lead list into personalized, legally-compliant cold-outreach emails — governed end to end by a structured prompt system rather than a one-off prompt. Built for [Ready Konnect Inc.](https://www.readykonnect.ca), a Southern Ontario staffing agency, consuming leads produced by [`gta-manufacturing-leads-scraper`](https://github.com/saintPadi/gta-manufacturing-leads-scraper).

This repo is less "a script that calls an LLM" and more a small case study in **prompt engineering as a discipline**: a reusable prompt library, explicit guardrails, and an automated QA pass that checks the model's own output before it's allowed to ship.

Why this is here

Most AI demos stop at "the prompt works once." The interesting engineering problem is what happens at draft #24 of a batch, after the model has settled into a repetitive cadence, or when a legal requirement (Canada's Anti-Spam Legislation) has to be satisfied in *every single output*, not just the one you eyeballed. This project's answer to that is three layered artifacts:

1. A reusable prompt library (reference/email_playbook.md) — not a single prompt, but a versioned set of rules: structure, tone, personalization inputs, subject-line rotation, and a CASL-compliant footer template. SKILL.md is the orchestrating system prompt that tells Claude when and how to apply it.
2. Guardrails baked into the prompt, not bolted on after — explicit "do not fabricate," "do not promise unverified numbers," and "never send, draft only" rules live in the prompt library itself, so the model can't accidentally reason its way around them.
3. Automated AI governance (scripts/check_humanizer.py) — a deterministic lint pass that runs after generation and flags AI "tells" (em dashes, stock cold-email phrases, run-on sentences, repeated subject lines, repeated paragraph openers) before a batch goes to a human for final review. It was a