Standardizing Chaos

Quick note: if you replied to a previous newsletter and it bounced, I'm sorry. I had some DNS issues when setting up Buttondown, and replies weren't routing correctly. It's fixed now. Hit reply anytime.

I've been deep in standardization work this week, mapping out how our services handle API specs. We have a lot of services. Everyone does things a little differently. Some teams generate specs from code. Some write specs first. Some do a mix of both.

I used Claude Code to search through all of our org's repos, filter for the ones using Fastify, then categorize them by their approach: code-only generated specs, mixed spec generation, or API-first with hand coded specs. That gave me a migration plan for each category.

Here's what keeps running through my head: if you don't have standards, you standardize chaos.

Working through that chaos is necessary. But you need a champion to lead it. The general group can't do it themselves, at least not yet. The standards need to be simple enough that engineers recognize them instantly. Copy-paste-able. People won't always do the "right" thing. So you need someone guiding the process until it becomes muscle memory.

This Week on Slightly Caffeinated

Ethiopian Coffee, Bun, and Building a Life Companion on Prism

TJ has been building an AI life companion using Prism. The system uses memory and conversation history to provide meaningful interactions. It tracks context, offers encouragement, and even told him it was proud of him after a rough day. We talked through the tech stack (Laravel, Prism, PGVector, Vercel's AI SDK), memory management approaches, and plans for MCP integrations like calendar and task management.

The episode drops tomorrow (Saturday) at 8:30am Eastern.

What I'm Learning

I've been using Claude Code for research and investigation at work. Specifically: scanning hundreds of repos to understand how teams approach API development.

The workflow: Claude Code searched our org's repos for Fastify dependencies, then I identified patterns to categorize them. Code-only specs, mixed generation, or API-first. From there, it researched each service and sorted them into buckets.

The result: a clear migration plan per category instead of a vague "we should standardize" statement. AI is good at tedious research across large codebases. The boring stuff that nobody wants to do manually. This allows me to look at organized data to decide what to do with it.

Links Worth Your Time

Writing a Good claude.md - Practical tips for creating claude/agents files. The advice on progressive disclosure stood out: don't front-load all your context. Link to docs instead of inlining them. Keep the core file small and let Claude pull in what it needs. I have some trimming to do on my own files.

How AI Is Transforming Work at Anthropic - Anthropic surveyed 132 of their own engineers about AI usage. The findings: engineers report using Claude in 60% of their work with a 50% productivity boost. 27% of Claude-assisted work wouldn't have been done otherwise (papercuts, nice-to-haves, exploratory work). The interesting tension: people are becoming more full-stack, but some worry about skill atrophy for deeper technical work. Worth reading the whole thing.

10 Holiday-Themed Kids AI Activities - This podcast episode walks through holiday activities you can do with kids using AI: generating stories, coloring pages, and more. I'm planning to try some of these over the Christmas and New Year's break.

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

-Chris