AI Cold Email Personalizer — Research Every Prospect and Write Genuine Outreach in 90 Seconds Without the Usual AI Fluff
Built by ApexHelix ·
About this build
A Claude and Perplexity powered workflow that researches each prospect individually before writing a single word — pulling their recent LinkedIn activity, company news, job postings and public content — then generates a cold email that opens with something real and specific to that person rather than a template that co
What was built
Built this after noticing a pattern across every cold email course and tool I had tried. They all optimised for the writing step and completely ignored the research step. The writing is the easy part. Knowing something real and specific about the person you are emailing is the hard part — and it is the only thing that actually makes cold email work. Here is exactly how the workflow runs: **Step 1 — Contact input** A CSV of prospects gets uploaded to Clay or imported from Apollo. Each row contains the prospect's name, job title, company name, LinkedIn URL and website. That is all the workflow needs to start. **Step 2 — Prospect research via Perplexity** For each prospect Perplexity runs 4 targeted searches simultaneously. Recent LinkedIn activity — any posts, comments or articles the prospect has published or engaged with in the last 60 days. This is the single most valuable research signal because it tells you what is actually on their mind right now not what their job description says they care about. Company news — any funding announcements, product launches, leadership changes, press mentions or notable events from the last 90 days for their company. Recent events create natural conversation openers that feel timely rather than generic. Job postings — what roles is their company currently hiring for. A company hiring 3 sales engineers is signalling something very different to a company hiring 3 customer success managers. Job postings reveal strategic priorities that have not been announced publicly yet. Personal mentions — any conference talks, podcast appearances, published articles or interviews featuring the prospect personally. Being referenced for something they created or said publicly is one of the most effective openers in cold email. **Step 3 — Research synthesis via Claude** All four research outputs get passed to Claude with a structured prompt that does three things. Selects the single strongest signal from everything found — the one piece of information most likely to be genuinely relevant to why someone would buy the product being sold. Not the most impressive signal. The most relevant one. Identifies the connection between that signal and the specific problem the product solves. This is the bridge step that most AI email tools skip entirely. Finding something interesting about a prospect is meaningless unless it connects to a reason they might care about the product. Flags any red flags — signals that suggest the prospect is not actually a good fit right now such as a recent layoff announcement, a competitor product launch that they just invested in, or a leadership change that might mean delayed decisions. **Step 4 — Email generation via Claude** Only prospects who clear the red flag check move to email generation. Claude writes a personalised email using a strict structure. Line 1 — opens with the specific research finding. Not a compliment. Not a generic observation. The actual thing found — a specific post they wrote, a specific announcement, a specific job listing. Under 25 words. Line 2 — connects that finding to the problem the product solves. One sentence. No product mention yet. Line 3 — introduces the product in one sentence focused on outcome not features. Line 4 — social proof in one sentence. Specific customer or result not a vague claim. Line 5 — call to action. One specific ask. Not let me know your thoughts. A specific question or a specific time offer. The entire email is under 100 words. Every word is there because it earns its place. **Step 5 — Output and review** Completed emails get written to Airtable with the prospect details, the research signal used, the email text and a confidence score Claude assigns based on how strong the research signal was. Low confidence scores go into a manual review queue rather than sending automatically. High confidence emails get pushed to Gmail or Instantly for sending. **Why the research-first approach changes everything:** The open rate difference between a generic AI email and a research-first email is not marginal. In testing across 400 prospects over 8 weeks the research-first approach generated a 34 percent open rate versus 18 percent for the template approach. Reply rate was 8.2 percent versus 2.1 percent. The more interesting finding was in the reply content. Generic AI emails that do get replies tend to get one-line dismissals. Research-first emails that get replies tend to get actual responses — people engaging with the specific thing referenced in the opening line. That engagement quality is what converts into booked meetings not just replies. **What broke during development:** The LinkedIn research was the most unreliable step early on. Perplexity does not always surface recent LinkedIn activity depending on the account's privacy settings and posting frequency. Fixed by adding a fallback instruction — if no LinkedIn activity is found in the last 60 days, fall back to company news. If no company news, fall back to personal mentions. If none of the above, flag the prospect for manual research rather than generating a low-quality email. The email length kept creeping up in early versions. Claude would add context and qualifications that made the emails feel thorough but killed response rates. Fixed with a hard word count instruction — 100 words maximum, no exceptions, and an explicit list of sentence types that are banned — preamble sentences, capability sentences and future-state sentences that describe what could happen rather than what has happened. **Results after 8 weeks:** 400 prospects contacted. 34 percent open rate. 8.2 percent reply rate. 23 booked meetings. 4 closed deals directly attributable to sequences started by this workflow. Average time from contact input to sent email — 94 seconds per prospect. The 4 closed deals all had one thing in common — the opening line referenced something the prospect had posted or announced in the previous 30 days. Recent and specific is the formula. Will share the full n8n workflow JSON, Perplexity research prompts and Claude email generation prompt in comments for anyone who wants to deploy this.