How I use AI for LinkedIn outreach (probably obvious to some of you but I keep seeing people mess this up)

A viral post from an AI practitioner is challenging conventional wisdom on using LLMs for LinkedIn outreach. The core argument is that most people misuse models like GPT-4 or Claude as a "creative layer," leading to generic, detectable AI text. Instead, the author advocates treating the LLM as a "pattern completer" and using hard constraints in the system prompt—such as strict word limits or mandatory reference to a specific trigger—to tightly control output. This method proves more reliable than soft style guidance like "sound conversational," which degrades over long contexts.

The author's operational system involves a multi-prompt architecture, where each message in a sequence has its own dedicated prompt and schema. The key to personalization is a robust "variable layer" populated with specific behavioral signals per lead, like recent job changes or inferred tech stack from postings. These real data points are slotted into named placeholders, giving the model the simpler task of rendering facts into sentences rather than inventing personalization from scratch. While effective, the manual pipeline (using Google Sheets and Python scripts) highlights a scaling challenge with subtle output repetition, pointing to an unmet need for reliable, automated agent-based enrichment and generation pipelines.