Welcome to the 51st issue of AI Agents Simplified 🍻
This issue is brought to you by Novu Connect
Today's newsletter is a special one! We are thrilled to host Sam Keen as our guest author for this issue. Beyond being a bestselling author, Sam brings massive experience from his time at AWS, Lululemon, and Nike. Check out his LinkedIn profile here to learn more about his work:
Open a fresh Claude Code session and the assistant knows nothing about your project: not where the tests live, not the patterns this codebase uses, not the conventions you walked it through yesterday. So you walk it through them again. And again tomorrow.
That re-teaching is a tax, and it’s the easiest one to overlook. It never shows up on a gauge, so you pay it at the start of every session without ever counting it.
The limiting factor in AI-assisted development right now is not how smart the model is. The models are plenty smart. The limiting factor is that none of that intelligence carries from one session to the next, and supplying the missing context has been left to you.
Picture a senior engineer who forgot everything about your codebase overnight, every night. Brilliant, fast, genuinely helpful between 9am and 6pm, and starting from zero again every morning. You would not call that person a force multiplier. You’d call it exhausting. Yet that’s the default relationship many people have with their coding assistant. They blame the model for the drag, when the real problem is that nothing they explained yesterday is still there today.
The fix is to stop starting from scratch. Write the recurring context down once, in a place the harness reads automatically, so it’s there in every session without you lifting a finger. Claude Code gives you several mechanisms for this. Three are foundational, and the useful way to think about them is by what kind of cost each one eliminates:
Agent files (CLAUDE.md) are your project’s standing memory: the conventions, the layout, the “we do it this way here.” Write them once and they load every session, so you stop re-explaining the project. They also load hierarchically, and that’s the underused part: a personal file that rides along on every session on your machine, a broader one scoped to all your coding work, and a project-specific one, each layering on top of the last. Keep them lean, a couple hundred lines each.
Skills capture the procedures you’d otherwise walk through by hand each time: the multi-step moves, the “first do this, then check that.” A procedure you’d re-explain becomes a procedure you invoke. And a skill isn’t limited to instructions: it can bundle scripts the agent runs, so the repeatable move can carry real code, not prose alone.
Hooks handle the corrections you’d otherwise repeat: the lint nit, the formatting rule, the check you keep asking for. They fire at fixed points in the harness lifecycle (before or after a tool runs, for instance) and sit outside the model’s control, so they run every time, whether or not the model would have remembered. You give the note once, and never again.
The pattern underneath all three is the same. Each one converts a recurring cost into a one-time write. That is what “compounds” means here, and it’s the opposite of starting from scratch. The investment is small and the payback lands in every session that follows.
If your Claude agent is trapped inside your app, it’s time to move the conversation to where your customers actually live. Without the headache of maintaining custom infrastructure.
Novu Connect is the new standard in Agent Communication Infrastructure (ACI). It bridges the gap between your AI agents and any messaging platform, including Slack, Teams, WhatsApp, and email. You don’t need to spend weeks building pipelines; just pick a template, connect your tools, and you’re live in under 2 minutes.
No infrastructure to babysit, no custom plumbing to rebuild. Just immediate, two-way conversations with your customers.
Join the challenge: deploy your agent in under 2 minutes and unlock a free month of our Pro tier.
There’s an honest catch, and this is where /context earns its keep. A compounding system can rot. You install a skill pack and forget it. A CLAUDE.md that started tight slowly fills with stuff that mattered once. The investment quietly turns into freight, and freight is just the from-scratch tax wearing a nicer costume.
/context is how you tell the difference. It does one useful thing: it makes the invisible visible. One command, a colored grid, a per-category breakdown of exactly what you’re carrying before the conversation even starts. You don’t need to master it. You need to glance at it.
When I last ran it, the single biggest chunk of my standing context wasn’t the project memory I’d carefully written. It was 14.4k tokens of skill packs I’d installed on a whim and never used once. I’d assumed my context was working for me, and a five-second look said otherwise. Culling them took a minute: /skills to deactivate the ones I don’t use, /plugin to drop a whole pack that arrived with a plugin.
The same readout flagged a second outlier with a different fix: the CLAUDE.md in my working directory had grown to 3,500 tokens. I sat down with Claude and compacted it, cutting the irrelevant and the quietly duplicated, and it came back at 1,200. Same project memory, a third of the weight.
Clarity helps the agent for the same reason it helps a human reader. A bloated, half-contradictory CLAUDE.md doesn’t only cost tokens. It muddies the very instructions you’re leaning on. A lean one is easier for the model to follow, the same way a tight brief is easier for a colleague to follow.
The specific number doesn’t matter. The habit does. Skim /context the way you’d skim a credit-card statement: not obsessively, only often enough to catch the recurring charge you forgot you signed up for.
Starting from scratch feels free because the cost is smeared thin across every session you’ll ever run. It isn’t free. It’s one of the largest, quietest line items in how you work.
So put in the upfront work: a lean memory file, a few skills you actually use, a couple of hooks that hold the line. Then keep curating. Models change, your projects change, and a glance at /context now and then tells you what’s still earning its place. The goal isn’t a clever prompt. It’s a setup that knows your project before you say a word.
If you want the hands-on version of this, it’s the spine of the live workshop I’m running with Packt on June 20: Claude Code Beyond Prompts. CLAUDE.md, skills, and hooks, built into a system that compounds.
Use Code : CLAUDE60 to get 60% off.
No posts
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.