Field notes
What breaks, why it breaks, and what to do about it.
Belief Shift
The Agentic Coworker Pattern - Giving a Claude Agent a Durable Role, Memory, and a Seat at the Standup
An AI coworker is not a chatbot that answers questions. It is an agent with a clear role, persistent memory of its own work, and a real slot in the team's cadence. Most teams build the first and skip the last two.
Read the note →
Apr 27, 2026
Framework + Case
Building a Custom Claude Agent: When the Anthropic SDK Beats LangChain and LangGraph
LangChain is the default. LangGraph is the upgrade path. Both add abstraction you pay for in production. Here is when the Anthropic SDK is the shorter build, the cheaper runtime, and the faster debug loop for a Claude-only agent.
Read the note →
Apr 26, 2026
Framework + Case
Claude Skills vs Tools vs MCP - Which Abstraction to Reach For
Claude ships three overlapping ways to extend what an agent can do: skills, custom tools, and MCP servers. They solve different problems and most teams pick the wrong one first.
Read the note →
Apr 26, 2026
Belief Shift
A Slack-to-Claude-Code Bridge That Runs 16 Cron Heartbeats Across My Brand Fleet
An AI agent that posts blockers to Slack you cannot action from your phone is not solving the problem, it is just relocating it. I rebuilt mine around a shared blocker queue and a cockpit skill that drains the queue in 2 minutes flat, not 2 hours of thread-hunting.
Read the note →
Apr 26, 2026
Framework + Case
The Ahrefs + Google Ads MCP Server That Saved a Content Agency 52 Hours a Week
A content agency was spending 3 hours per client per week on Ahrefs audits. An MCP server turned those 3 hours into 20 minutes per account and unblocked 5 more AI use cases in the process.
Read the note →
Apr 25, 2026
Belief Shift
Claude Code as a Team Force Multiplier - What Changes When Every Operator Has an AI Coworker With Repo Access
Claude Code is not Cursor with a bigger context window. It is the first AI product that ships durable team-level leverage instead of per-seat chat speedup. Here is what actually changes when a team adopts it.
Read the note →
Apr 25, 2026
Depth Charge
Why Cold Email Stops Working at 50 Inboxes (And What to Do Instead)
Scaling cold email past the low hundreds per day looks like a tooling upgrade. It is a deliverability architecture problem and the usual fix makes it worse.
Read the note →
Apr 25, 2026
Contrarian
AI Resume Screening Is Filtering Out Your Best Candidates. Here Is What Replaces It.
AI resume screening tools are sold as efficiency. The published research shows they reject 27 million qualified candidates in the United States alone. The fix is not a smarter screener. It is a different architecture.
Read the note →
Apr 24, 2026
Contrarian
The Best CRM for a Small Business in 2026 Depends on One Thing You Haven't Checked
Every CRM comparison ranks features. None of them check whether the API at the tier you can afford will let your automation layer actually do anything. Here is the shortlist that does.
Read the note →
Apr 24, 2026
Framework + Case
Your First Custom MCP Server - When It Is Worth Building vs Using an Off-the-Shelf One
The official Anthropic MCP registry has servers for Slack, GitHub, Google Drive, and dozens more. Custom only makes sense in specific cases. Here are the 5 signals that custom is the right call.
Read the note →
Apr 24, 2026
Framework + Case
MCP vs REST APIs vs Function Calling - Which Abstraction to Reach For
Three abstractions expose data and actions to an LLM. They look interchangeable. They are not. Here is the decision tree.
Read the note →
Apr 24, 2026
Framework + Case
The n8n Architecture Pattern That Survives a 10x Volume Spike
Most n8n workflows that work fine at 100 executions a day fall over at 1,000. Not because of n8n but because of how they were wired. Here is the pattern that survives.
Read the note →
Apr 23, 2026
Framework + Case
Self-Hosted n8n vs n8n Cloud - The Cost Breakdown at 1K, 10K, and 100K Executions
The self-hosted n8n story sounds great until you count ops time. Here is the real cost comparison at 1K, 10K, and 100K executions per month, including the hour-a-month you actually need for maintenance.
Read the note →
Apr 23, 2026
Framework + Case
AI Trade-In Valuation - How Independent Dealers Can Beat CarMax on Speed Without Overpaying
CarMax's instant offer is not a price advantage. It is a speed advantage. Customers leave because they can get a 90-second number from CarMax while your appraiser is on another lot. AI valuation closes that gap without forcing you to overpay.
Read the note →
Apr 22, 2026
Framework + Case
n8n vs Zapier vs Make in 2026 - When Each Actually Wins
The automation platform war hit a new equilibrium in 2026. n8n matured into a serious self-hostable option. Zapier added AI-first features at the top of its pricing curve. Make became the pragmatic middle. Here is where each actually wins.
Read the note →
Apr 22, 2026
Framework + Case
DMS Integration Without the $15,000 Setup Fee - What n8n and Your CSV Export Can Actually Do
Franchise AI tools quote $5K to $15K for DMS integration. Most of what they are selling is a scheduled CSV import with error handling. n8n does that for free, and it works with every major independent DMS.
Read the note →
Apr 21, 2026
Framework + Case
Lead Response in 60 Seconds - The Cheapest AI Build That Moves a Dealership's Close Rate
Response time is the single most-studied input to lead conversion. A 5-minute response is 9x more likely to convert than 30 minutes. A 60-second response, if you can build it, is the highest-leverage AI project a dealer can ship first.
Read the note →
Apr 21, 2026
Framework + Case
How Independent Used-Car Dealers Are Replacing Their BDC With AI That Responds in 60 Seconds
A human BDC costs $35,000 to $60,000 per year per seat and responds in 45 to 90 minutes on average. An AI-first BDC responds in 60 seconds and does not sleep. Here is exactly what to build.
Read the note →
Apr 20, 2026
Belief Shift
Why the Big Dealership AI Tools Do Not Fit 30-Vehicle Lots
Numa, Toma, Matador AI, and DealerAI are priced and built for franchise dealer groups with 100+ vehicles per rooftop. They work. They are not wrong. They are just a bad fit for a 30-unit independent lot.
Read the note →
Apr 20, 2026
Framework + Case
How Independent Dealers Use AI to Surface Auction and Trade-In Inventory Without a Buyer
Inventory acquisition is the slowest lever at an independent lot. Manheim and ADESA list thousands of vehicles a week. An AI-first acquisition layer surfaces the right 5 to 10 for your lot in 15 minutes per auction day.
Read the note →
Apr 20, 2026
Belief Shift
You Don't Have a Tool Problem. You Have an Architecture Problem.
Most operators diagnose their AI pain as a tool problem. It is not. Adding another SaaS to a broken architecture just makes the leak bigger.
Read the note →
Apr 18, 2026