I've been running cold email campaigns for clients for 3 years and the biggest shift I've seen isn't the tools. It's what actually gets a reply. Personalization used to mean scraping a name and company from LinkedIn, dropping it in the first line, and hitting send. "Hey {FirstName}, I noticed {Company} and thought..." That worked in 2022. It's dead now. Everyone's doing it and prospects can spot a mail merge from the subject line. What changed for me was treating personalization like actual research instead of a data field. Here's what I started doing: → I scrape the prospect's entire website. Not just the homepage. Blog posts, service pages, case studies, about page, even their contact form if it's there. → Then I feed all of that into OpenAI and have it analyze what they actually do, who they serve, and what problems they're likely dealing with. The AI doesn't just summarize. It finds the specific details nobody mentions in generic outreach. So instead of "I saw you work in logistics," the email opens with "Noticed you handle cross border freight into Mexico. Your blog mentioned customs delays eating 15% of delivery windows." That's the kind of line that gets opened because it doesn't sound like 500 other emails they got that week. The reply rates went from 2-3% with generic personalization to 8-10% with actual research. One prospect replied last week: "Your email won because you actually read our site. Everyone else sent the same template." The system I built does this automatically. Scrapes the website. Analyzes every page. Generates icebreakers that reference non-obvious details. It writes openers like a human who spent 20 minutes studying their business, except it does it for 1,000 prospects in an hour. Here's what I learned building this: Small prompt details make a massive difference. Having OpenAI shorten company names naturally (say "Stripe" not "Stripe Inc.") and reference specific pages beyond the homepage makes it feel real. The difference between "I saw your website" and "I saw your freight tracking dashboard lets customers get ETAs without calling" is everything.

Am I trying to reinvent the wheel for my first ICM project? I'm working on building job search tools/workflows to support my own job search. I've jumped in head first and: - I'm burning through my CLAUDE PRO tokens - why? My sense is I've setup my root CLAUDE.md file poorly. I switched from Sonnet to Haiku and never used Opus. (I know there are other ways to access LLMs, but I want to spend some time mastering the Claude Code extension interface before jumping into openrouter, Claude API, etc.) - I'm iterating the folder/workflow/skill design but still getting really poor draft outputs (I got better outputs just prompting in the free version of Claude) - I'm still learning the Classroom content and am hoping to get some insights as to any red flags for being newer to this (i.e. I'm just biting off too much for my first project) ## Workspace Structure ``` jobbot/ ├── CLAUDE.md ← You are here (always loaded) ├── CONTEXT.md ← Task router ├── _config/ │ └── story-bank.md ← Master STAR narrative library (shared across workspaces) │ ├── linkedin-optimizer/ ← Build and maintain LinkedIn presence + content │ ├── CLAUDE.md │ ├── CONTEXT.md │ ├── profile/ ← Profile optimization pipeline │ ├── content/ ← Content creation pipeline │ ├── growth/ ← Analytics and growth tracking │ └── _config/ ← Stable reference: profile, voice, keywords, pillars │ ├── resume-optimizer/ ← Master CV + JD-tailored resumes │ ├── CLAUDE.md │ ├── CONTEXT.md │ ├── master-cv/ ← Full CV maintenance pipeline │ ├── tailored/ ← Per-application tailored resume pipeline │ └── _skills/ ← 6 skills: audit, rewrite, jd-decoder, match, tailor, ats-check │ ├── cover-letter-optimizer/ ← Per-application cover letters │ ├── CLAUDE.md │ ├── CONTEXT.md │ ├── pipeline/ ← intel → narrative → draft → review │ ├── final/ ← Approved cover letters │ └── _skills/ ← 4 skills: company-intel, narrative-select, draft, review │ ├── application-engine/ ← Application prep, ATS forms, tracking, analytics

One thing I’ve found is there’s no time for true work during the week. So I find myself binging this community, working on side projects, and building out my architecture on the weekends. Would love to hear about some projects you folks are working on! What’s going down this weekend?

It’s not: “we cleaned files” It’s: “we built a system that keeps itself clean.” The first month I was just trying to get my head around the Foundation, Implementation and understand what is happening. Foundation already gave me a Content Provider style architecture, along with the animation and website builder, i intuitively and with both ChatGPT and Claude I started by implementing Jake's structure. I watched YouTube videos as well that gave me other insights (and validated that Jake's course is the only place that i see that talks about ICM and Architecture and such). Either way, I signed up for premium yesterday, and found that there was much more to complement the Foundation. As Jake suggested, I slowly took each on of the Vault items, and explored with Claude - how is our architecture compared to GitHub repos (which I tapped into some of the members GitHub repos and made Claude check itself against those insights - super thank you Community!!). It pointed out a few times that our architecture is more mature, though it always found 1 or 2 items that could improve its flow. We had a glitch this morning when I prompted the execution of today's schedule with all the items to post, create and monitor, which the Orchestrator and Operator (me) had to dissect what it did and ensure it won't follow that course again, and whether such a flow is efficient both Token costs and Workflow. It found ways to optimize - extracted what it found useful from the Vault, informed that skills that mention in Vault are skills we already captured (among others). My very first project with Jake - text to video animation has been improving on its own outside my Brand Sandbox. And as I closed my Claude after uploading the content it had structured neatly for me in an HTML file with EVERYTHING that i needed for the content posting, I had it summarize its work - which I hope this help if you read this far (thank you for reading my stuff:)) - Three-day architecture delta (Apr 30 → May 3)

Found this great post on Substack by Ilia Katelin on how Claude operates under the hood and how it’ll processes tokens. The post starts with very basic prompting strategies then goes deeper into to settings and eventually on how to configure Claude files. We learn many of these in this community and through the ICM system, but there are some great nuggets here. One of my big takeaways for me is that Claude uses less tokens to ingest a document via a connector (eg Google Drive) vs dropping it into the context window itself. 🤯 Hope this brings value! update: I let Notebook LM do its magic and also included a PDF version of the full article in case anyone has trouble with Substack.