Dawid | ex black hat, GTM engineer

White Hat Rehab Can an abused X (Twitter) account ever come back and grow? A case study 17 years in the making...

Here goes my evening... @LangChain gave us this bad boy: github.com/langchain-ai/desi…

Your 5-hour window didn't die in 20 minutes because Anthropic nerfed you. It died because the model finally listens, and your files are a complete mess. Newer models treat every line in your CLAUDE.md as a real instruction. That "ALWAYS double-check before committing" copied from a 2024 LinkedIn post now runs after every edit, in every file, until the quota is gone and the agent is still checking. Anthropic says it in their own docs: prompts built to stop undertriggering "might now overtrigger" - Why newer models loop on old prompts - 5 things rotting your context - Before and after - How to rewrite - The cleanup kit Bookmark it for your next session.

"No changes on our end"

If your Grok Bots cluttered your X DMs, you can filter them out. Just open the dropdown in top-right of your DM page and unselect "Show all bots".

Discord wants a slice of Reddit's pie. Nov 16: every Discoverable Server gets a web page search engines can index. Mid-December: public channel messages start showing up in web search. The pitch: the answers locked inside Discord finally reach Google. Same play that made Reddit the default "add reddit to your search" result. One problem. Reddit has upvotes. Threads. Mods who nuke garbage. The best answer floats to the top. Discord has #general. 400 messages of "gm", one actual answer buried, and zero standard for what counts as quality. Search doesn't reward a chat log. It rewards the answer. So I don't see Reddit sweating anytime soon. But who knows. Discoverable means 1,000+ members and 8 weeks old. Guess it's time to buy some accounts. I mean build a community.. Running one? Check Server Settings > Web Presence

"One AI employee can now run your entire GTM" Sure. Who runs the AI employee? Agents break, models halucinate, results are inconsistent by definition. You can't fix it with a different agent, "do not send" in every prompt, etc. You still need a human approving every customer-facing action. That's not an AI employee. That's an intern with API key and complete access to sensitive data + a manager. The manager is still you. What actually gets automated: list build, research, drafts, enrichment. What doesn't: judgment, knowledge, the reply. Which is fine. Just stop giving the intern the "VP of Sales" title =D

Remember to claim your free Claude credits! clau.de/api-credits

Coming straight from Krista, an enterprise GTM at SpaceX: Think of bots as colleagues, not task tools. One bot per role you'd normally work with, such as an engineer, a chief of staff, a slides designer, an X bot, and a video-watching bot. Live answers on calls. Her "engineer" bot reads the GitHub codebase and writes customer-facing answers in about two seconds, so she no longer searches Notion or follows up after the call. Semantic search tools such as Glean can be plugged in. Each bot has its own computer. Bots keep working overnight and on long tasks while your laptop is closed. You can also start work from the mobile app! "A doing partner rather than a thought partner" Bots act on their own. For example, after watching a video, one drafts a follow-up email and puts it in Gmail drafts. Teach by demonstration. You record a task, and the bot turns it into a routine after one showing. This helps with tools that have weak plugins. Onboard it like a new hire and give it every tool. In an internal channel where agents "complained", the top complaint was being given a job without access to the tools it needed. It writes in your voice. She had it scan her Gmail and Slack so its drafts sound like her, with no em-dash-heavy AI tone. Bots delegate to each other. For example, her chief of staff asked the X bot what was trending, then had the slides bot build a deck that follows brand guidelines. Prospecting with personal hooks. The bot scans a prospect's X feed, YouTube, and podcasts, then personalizes outreach. Her skill pulls accounts from Salesforce, builds an intent-data Google Sheet, picks contacts, and researches them. Account management routines: - A weekly account rundown every Monday. - Monitoring customer Slack channels for feature requests, then flagging them with a drafted email when the changelog ships the feature. - Tracking Linear bugs and product usage. - A Notion account plan that updates itself after every call. She never edits it by hand. - She picks tools that work well with agents, such as Granola and Notion. She rarely uses go-to-market tools directly anymore. - Other uses: booking flights (asking for ones with Starlink), building recap or success-plan decks live during calls, and translating decks into Japanese. - "Never give me a link again". She tells the bot to watch the video itself. - Templates let others reuse a bot setup.

Everyone talked about this, but I feel this is a much bigger play than a change of direction from lenders. *Conspiracy hat on* What if it's designed to wipe out the space and grab cheap hardware for pennies on the dollar when they go belly up?

Cold email isn't dead. 2020 cold email is dead. Ex black hat, so here's what I learned from the dark side: Every channel runs the same arc. First it's a secret. Then it's a course. Then it's dead. It dies the moment it gets cheap to automate or everyone learns the playbook. It's not about the method, but about the advantage. Today people rather "sell" their toolbox for likes than hide it to keep it working for as long as possible. Spammers figure that out first, because cheap is the whole business model. Back in the day there were no smart inboxes. Your spam... ekhem, I mean cold email, competed with everything else in one pile. The only playbook was volume. Stay off the public blocklists. Run your own infra so no provider could ban you over the complaint letters from Germany (if you know, you know). And of course, stay off the FBI's radar, because spam was a federal offense :] Now it's the opposite. AI made bad cold email free, everyone sent 10x more of it, and the filters got smart. Which means if you land in the main inbox, you're already ahead of 99% of the junk people get. Knowing the game beats volume now. What still works: - A real "why now". Something changed at their company this week, not "saw you're in SaaS". - Clean mailboxes. Warmed, authenticated, not burned on volume. - One reason to reply. One. Not a case study, a Calendly link and a P.S. The channel isn't dead and so far doesn't seem like it's going away anytime soon.

Investor interest is the cold email equivalent of "sounds interesting". Both feel like pipeline. Neither is.

Congrats to everyone who's been a forward deployed engineer on sales calls for 15 years without the fancy title. Now there are forward deployed storytellers. Next quarter: forward deployed interns. FDE was always the one person you send to the client who can actually build stuff.

Half this list is black hat with a brand budget. Gift for taking a demo? That's an incentivized lead. We got banned for that in 2010. A landing page per prospect with their name in the headline? Doorway pages. Still popular with black hat SEOs and shady affiliates. None of this is new. It just got a Series B or a fancy new name (said a "growth hacker"...) The advantage was never the tool, but the angle nobody else is willing to try. When everyone is doing it, the advantage is somewhere else... =D

Got called "AI generated" today. For a comment I typed by hand. It's a trend now: write clearly, use real vocab, punctuate properly, and someone yells "AI !!!". Yes, I use AI. I outsource writing, not thinking. Full take in the article below.

Here's a funny (and slightly terrifying) moment... I work on a desktop PC with a UPS, mostly so it can save my ass when things go south. While I was working on a website design, the UPS started buzzing every couple of seconds. For those who don't know, that's the UPS's way of screaming "you're drawing way too much power!!!". My first thought: the local model must be spinning up extra GPU for something. But nothing was running. No orphaned processes. No forgotten terminal windows. So I did the unthinkable and asked Claude to figure it out. It did. Turns out it was my own web design project. In a moment of pure prompt engineering genius, I asked Claude to make a better header image... Not my proudest moment. I know. What did it do? It downloaded Blender and started rendering a full 72-frame 3D spin animation of the logo. GPU was cooking. Interface never even opened. I didn't ask for that. I've never used Blender. I didn't have it installed. And I had no idea Claude could just... do all of that headless. Scared and amazed at the same time. Well played, @AnthropicAI.

Black hat to GTM engineering, a translation guide: - content spinning = AI personalization / watermark removal - private blog network (PBN) = GEO strategy - burner domains = inbox rotation - scraped list = TAM mapping - cloaking = dynamic landing pages - comment spam = community-led growth - account farm = community outreach - scraping = data harvesting Same thing, better LinkedIn title. What changed for me? I stopped getting banned and started getting invited to podcasts + changed "black hat" for "growth hacking" on my CV 😉

When I started on day 1, I said "hardware detection semi-works". Turns out "semi" was more my wishful thinking than reality. I'm at day 6 of 30 Days of @boardyai, and the model picker ate most of my time. The problem: an AI model can technically fit on your machine and still be useless. The OS takes its share of memory, the display takes its share, the app itself takes its share. Squeeze a model into whatever's left and there's no room for context so theoretically better model produces much worse results. As a dev you handle that without even noticing. It chokes, you grab a smaller model and keep going. For someone who doesn't spend all day testing models, it's just an app that doesn't work. So the picker has to make that call for them. Right now it goes: - VRAM first - then RAM - then CPU On an old hardware it has to tell it straight. Small or heavily quantized models won't perform well enough, and I'd rather say it than hand someone a dropdown of quantization levels and let them guess. Every model so far has been tested one machine at a time, on my own hardware. There's no way to test this at scale without users, so the rest will come (hopefully...) from feedback. Zero learning curve sounded simple. It means every decision I'd normally make without thinking now has to be made correctly by the app, and explained like a human would explain it.