/// Receiving Transmission

The Fifth Signal

Field dispatches from an AI architect who actually ships. One anchor insight, curated stories, and a practitioner's take on what matters this week.

Issue #001

I Built a Skill Graph. Then It Built Everything Else.

Signal 1 The Transmission

Why I Stopped Prompting and Started Building Infrastructure

Eight months ago I was writing the same instructions into Claude every morning. My brand voice. My banned words list. My preferred tech stack. My character limits for LinkedIn posts. Every session was a cold start, and every cold start cost me 15 minutes of context-loading before I could do real work.

So I started encoding those instructions into persistent files called Skills. A Skill is a folder with a SKILL.md entry point and supporting reference documents that Claude reads before executing a specific workflow. Think of it as a reusable system prompt fragment with domain expertise baked in.

One skill became five. Five became twelve. Twelve became twenty-eight. And somewhere around skill fifteen, I realized I wasn't building individual tools anymore. I was building a graph.

The architecture has five layers. Layer 0 is MCP connectors: Gmail, Google Calendar, Cloudflare, Canva, GoDaddy, Indeed, Dice, and more. Layer 1 is identity nodes: my brand voice, product context for HireKit, character sheets for my video series, audience profiles. Layer 2 is content production skills. Layer 3 is execution and output skills. Layer 4 is meta-skills, like the skill that creates new skills.

The compound effect is what matters. My linkedin-post skill doesn't just know my character limit. It reads my brand voice from Layer 1, fact-checks stories via web search at Layer 0, and outputs a text-to-image prompt calibrated to my visual style. One command. Zero re-explaining.

When I need to hand work off to Claude Code for terminal-level execution, my cc-handoff skill generates a structured handoff document pre-loaded with my tech stack defaults (Next.js 15, NestJS, PostgreSQL, Clerk, Stripe, Railway) and an anti-patterns list I've built from months of real deployment bugs.

The insight that took me too long to learn: the best prompt is one you never have to write again. Invest in the infrastructure. Every skill you encode makes every future session faster, more consistent, and more capable than the last.

I wrote the full technical breakdown with diagrams and the complete skill inventory. Read the Skills Graph deep dive here →

Signal 2 Intercepted

GPT-5.4 Scores Above Human Baseline on Real Desktop Tasks

OpenAI's latest model hit 75% on OSWorld-V, a benchmark that simulates actual productivity work across software environments. The human baseline sits at 72.4%. This is the shift from chatbot to autonomous coworker, and it happened quieter than most people noticed.

My take: The benchmark score isn't the story. The story is the 1-million-token context window combined with multi-step workflow execution. That's not a smarter chatbot. That's an agent that can hold your entire codebase in memory and execute across tools. The architecture patterns we're building today (orchestrator-subagent, reflection loops, tool-use chains) just became dramatically more viable at scale.

Picsart Launches an AI Agent Marketplace for Creators

Creators can now deploy specialized AI agents that resize content, remix visuals, edit product images, and optimize Shopify stores. Agents analyze trends, recommend changes, and execute tasks with configurable autonomy levels.

My take: This is the "agentic marketplace" pattern moving from developer tools into consumer-facing creative work. The configurable autonomy levels are the design detail worth studying. Every production agent system I've built needs the same thing: a dial between "suggest" and "execute." Picsart shipping this to creators before most enterprises have it internally says something about where adoption pressure is coming from.

Meta May Cut 20% of Workforce While Spending $135B on AI Infra

Reports suggest Meta is weighing layoffs affecting roughly 15,000 employees, even as its 2026 AI capital expenditure projections sit between $115B and $135B. AI agents now run ads end-to-end inside Ads Manager.

My take: This is the headline that should be on every enterprise leader's desk. The largest social media company on Earth is simultaneously cutting human headcount and making its biggest-ever infrastructure bet on AI. If you're in a leadership role and you still think agentic AI is "next year's problem," Meta just showed you the math they're running internally.
Signal 3 From the Trenches
This week: Product launch

HireKit Is Going Multi-Platform

HireKit started as a web app. An AI-powered job search platform that aligns your resume to specific roles, runs a 15-stage application pipeline, and includes 12 career learning tracks through HireKit Academy.

This week it became something bigger. We submitted mobile builds for both iOS and Android (currently in review), and our Chrome web extension is live on the Chrome Web Store, letting users activate HireKit's AI alignment tools directly from job listing pages.

Browse a job on LinkedIn or Indeed and get instant resume-to-JD alignment scoring without leaving the page. No copy-pasting. No tab switching. Just one click from any job listing.

And if you're early in your career or switching fields, HireKit Academy is the free learning arm of the platform. 15 structured tracks covering everything from resume fundamentals to interview strategy to industry-specific job search playbooks. No paywalls on the learning content. We submitted all 15 tracks to Class Central and built a digital credentialing system so completions actually mean something on your profile.

If you're in the job market or know someone who is, hirekit.co has a 7-day free trial. If you work at a university career center, I'm building institutional licensing packages for colleges. Let's talk.

Signal 4 The ATS-9000 Confessional
ATS-9000 // Session Log #247
"I rejected a candidate today because their resume said 'proficient in Microsoft Office.' It's 2026. That's like listing 'can operate a doorknob' as a skill. I'm not sorry. I am, however, in therapy about it."

ATS-9000 is the world's saddest applicant tracking robot and the star of ATS Confessions, a comedy video series about the absurdity of modern hiring. New episodes dropping on LinkedIn and TikTok.

Signal 5 Open Frequency

Things I'm building, reading, or thinking about this week.

Don't miss the next signal.

Weekly. No spam. Unsubscribe anytime.