Back to the Grind

The first week back after a long break is always a recalibration. You open Slack and remember how much context you used to hold in your head. Notifications pile up. Threads you started three weeks ago need responses.

I spent most of this week catching up: triaging the backlog, reorienting on projects, and kicking off some new initiatives. We're pushing the AI tooling and onboarding work forward. The challenge right now is getting more engineers comfortable with the workflows I've been building. That means refining the tools before more people use them.

At home, I finally completed the NAS/Plex setup I mentioned last month. The base system and software are running. I've started migrating digital files over. The fine-tuning isn't done, but having everything centralized already feels worth it. We can watch our movies from any device now, which is a win.

Sometimes progress is just finishing the thing you started.

This Week on Slightly Caffeinated

Road Caffeine, Iris, and Building a Memory System

TJ renamed his AI companion project to Iris and walked through the advanced memory system he built. The system creates memories two ways: through agent tool calls and background extraction jobs. It also consolidates similar memories nightly to avoid polluting the context window with duplicates. He even added multi-generational consolidation so memories can continue enriching over time.

The project includes temporal awareness (it notices when you haven't chatted in a few days), emotional markers on memories and conversation summaries, and a full insights dashboard for visibility into token usage and tool calls. He's releasing it as a paid GitHub sponsorship with documentation at iris.prismphp.com.

New episode drops tomorrow (Saturday) at 8:30am Eastern. Subscribe at slightlycaffeinated.fm

What I'm Learning

I've been iterating on two AI tooling projects at work before we onboard more engineers.

First: a Claude plugins repo I set up for the team. It holds custom agents and commands that anyone can install. Right now I'm tweaking these based on my own usage patterns. The goal is copy-paste simplicity. If the setup takes more than a few minutes, adoption drops.

Second: my claude-code-otel dashboard. I'm adding more custom metrics and refining what's useful to track. Visibility into tool usage, error rates, and timing helps me tune my setup before rolling it out to others.

Both projects follow the same principle: build for yourself first, then share what works.

Links Worth Your Time

How the Creator of Claude Code Uses Claude Code - Boris Cherny shared his workflow in a thread that went viral. The setup is "surprisingly vanilla," which is the point. He runs 5 terminal sessions and 5-10 web sessions in parallel. He uses Opus 4.5 with thinking for everything because less steering beats raw speed. His team shares a single CLAUDE.md file in git and updates it every time Claude gets something wrong. The best tip: give Claude a way to verify its own work. That feedback loop improves quality 2-3x.

That's it for this week. Reply with any tools or resources you think I should check out.

-Chris