{"slug": "i-built-an-ai-that-roasts-cold-emails-here-s-what-18000-drafts-taught-me", "title": "I Built an AI That Roasts Cold Emails — Here's What 18,000 Drafts Taught Me", "summary": "A developer built an AI tool called RoastMyEmail that scored over 18,000 cold email drafts and found the average first draft scores just 32 out of 100. The analysis revealed that most cold emails fail for the same systemic reasons: generic openers, buzzword-heavy value propositions, and overly aggressive calls to action like requesting a 30-minute call. The tool evaluates drafts across four dimensions and provides blunt, specific feedback to close the gap between what senders think they're communicating and what recipients actually experience.", "body_md": "Most cold emails deserve to be ignored.\n\nI know that sounds harsh. But after scoring 18,000+ cold email drafts through [RoastMyEmail](https://www.roastmyemail.fun?utm_source=devto&utm_medium=article&utm_campaign=launch), the tool I built and launched this year, the data is pretty unambiguous. The average score on a first draft is **32 out of 100**. That's not a rounding error. That's a systemic failure across industries, seniority levels, and company sizes.\n\nAnd the worst part? Almost every bad email fails for the exact same reasons.\n\nThat's what this post is about: why I built the tool, what the data showed, and what I'd tell anyone writing cold email right now.\n\nI've been on both sides of cold email. Sending it as a founder trying to get early customers. Receiving it as someone who runs a small product and apparently looks like a good prospect to every B2B SaaS company in existence.\n\nThe gap between what senders think they're sending and what recipients actually experience is enormous.\n\nSenders think their email is personalised because they swapped in a company name. Recipients see “I love what you're doing at Acme Corp” and immediately know it's a blast. Senders think their value prop is clear because they spent three hours writing it. Recipients read “leverage cutting-edge AI to drive synergy and scale revenue 10x” and have learned nothing. Senders think asking for a 30-minute call is reasonable. Recipients see it as asking for a marriage proposal on a first date.\n\nThe problem isn't effort. Most senders are trying. The problem is that **you cannot read your own cold email the way a stranger reads it.** You know your product, your intent, your proof. The stranger only knows what's on the screen. That blind spot is where reply rates die.\n\nI wanted to build something that closed that gap. A tool that reads your draft the way a stranger would, tells you exactly what's failing, and gives you a direction to fix it. Not a generic grammar checker. Not a “score” with no explanation. A brutally specific critique before the email reaches a real inbox.\n\nThat's RoastMyEmail.\n\nThe core loop is simple:\n\nThe AI evaluates four dimensions independently:\n\nEach gets its own score and specific feedback. Not a vague “could be better” but a callout like:\n\n“Your opener could go to 10,000 people without changing a word.”\n\nOr:\n\n“You've used three buzzwords in one sentence and communicated nothing concrete.”\n\nThe roasts are intentionally blunt. That's the point. Polite feedback on cold email is useless. A colleague who says “looks good to me” isn't reading it as a stranger. The AI doesn't pull punches because real prospects don't either.\n\nAfter scoring tens of thousands of cold emails, the patterns are depressingly consistent. Here are the findings that surprised me most:\n\nNot 60. Not 50. 32.\n\nMost cold email going out into the world right now is failing before a human even decides whether to read it.\n\nIt appears in roughly 1 in 4 drafts.\n\nIt signals bulk send before the recipient finishes the first sentence. It has never helped anyone get a reply. It persists entirely because senders default to it when they run out of things to say.\n\nSame problem.\n\nIt's a throat-clearing line that tells the recipient nothing about why this email exists. Cutting it entirely and starting with the actual point raises scores significantly.\n\nThis one is obvious in retrospect but striking in the data.\n\nEmails that use words like:\n\nin the first three sentences almost always have value props that can't be explained in plain English.\n\nThe buzzwords aren't decoration. They're a signal that the sender hasn't figured out what they're actually offering.\n\nAsking for a 30-minute call in email one is the most common CTA pattern.\n\nReplacing it with a yes/no question or a permission ask like:\n\n“Should I send the one-pager?”\n\nis the single change that moves scores the most on a rewrite.\n\nIt's also the easiest change to make. One line. No restructuring required.\n\nThey read like one person writing to one person.\n\nNot a campaign. Not a template. Not a pitch deck in email form.\n\nOne human who looked at another human's world and had something genuinely relevant to say about it.\n\nEverything else, including length, format, and subject line structure, matters less than that.\n\nThe scoring model was the obvious challenge.\n\nGetting an AI to evaluate “personalisation depth” and “hook quality” in a way that's consistent, specific, and actually useful, not just pattern-matching surface-level spam signals, took a lot of iteration.\n\nBut the harder problem was the feedback quality.\n\nA tool that tells you:\n\n“Your opener is generic.”\n\nis marginally better than nothing.\n\nA tool that tells you:\n\n“Your opener — ‘I wanted to reach out because I noticed your company is growing’ — signals bulk send, contains zero specific research, and could be sent to 10,000 people without changing a word.”\n\nis actually useful.\n\nGetting the feedback to be that specific, consistently, is what took the most work.\n\nThe score is the hook.\n\nThe line-by-line feedback is the product.\n\nThe other thing I didn't anticipate was the leaderboard.\n\nI added a public leaderboard of the worst-scoring emails, anonymised, mostly as a fun feature. It became one of the stickiest parts of the product.\n\nPeople love seeing how bad the worst emails are.\n\nIt also creates a natural benchmark. If your email scores 32, you can see exactly where that puts you.\n\nIf your first sentence could be sent to 1,000 people without changing a word, it is not a cold email opener.\n\nIt is a blast opener.\n\nFind one specific, verifiable thing about this recipient:\n\nWrite that as your first sentence.\n\nIt takes five minutes of research per prospect.\n\nIt is the highest-return change you can make.\n\nNot a feature list.\n\nNot a buzzword stack.\n\nThis is a value prop:\n\n“We helped [Company] reduce onboarding drop-off by 34% in one quarter.”\n\nThis is not:\n\n“We leverage cutting-edge AI to drive holistic revenue synergy.”\n\nThe test:\n\nCan a stranger read it and know exactly what happens if they buy?\n\nIf not, rewrite it.\n\nExamples:\n\nThese are low-friction asks that a busy person can answer in ten seconds.\n\nA 30-minute calendar hold from a complete stranger is a large ask that busy people decline by not replying.\n\nEmail one earns a conversation.\n\nIt does not earn a meeting.\n\nRoastMyEmail is free to use.\n\nPaste any cold email draft, subject line and body, and get a 0–100 score with line-by-line feedback in seconds.\n\n👉 [https://www.roastmyemail.fun](https://www.roastmyemail.fun)\n\nIf you're building something in the outbound, sales, or email space and want to talk shop, I'm always up for it.\n\nFind me at [https://x.com/roastmyemail](https://x.com/roastmyemail) or drop a comment below.\n\nAnd if your first draft scores above 70, genuinely impressive.\n\nMost people need at least one rewrite to get there.", "url": "https://wpnews.pro/news/i-built-an-ai-that-roasts-cold-emails-here-s-what-18000-drafts-taught-me", "canonical_source": "https://dev.to/ahsangadit96/i-built-an-ai-that-roasts-cold-emails-heres-what-18000-drafts-taught-me-3g4p", "published_at": "2026-05-28 12:16:31+00:00", "updated_at": "2026-05-28 12:22:23.235031+00:00", "lang": "en", "topics": ["ai-tools", "ai-startups", "ai-products", "artificial-intelligence", "natural-language-processing"], "entities": ["RoastMyEmail", "Acme Corp"], "alternates": {"html": "https://wpnews.pro/news/i-built-an-ai-that-roasts-cold-emails-here-s-what-18000-drafts-taught-me", "markdown": "https://wpnews.pro/news/i-built-an-ai-that-roasts-cold-emails-here-s-what-18000-drafts-taught-me.md", "text": "https://wpnews.pro/news/i-built-an-ai-that-roasts-cold-emails-here-s-what-18000-drafts-taught-me.txt", "jsonld": "https://wpnews.pro/news/i-built-an-ai-that-roasts-cold-emails-here-s-what-18000-drafts-taught-me.jsonld"}}