Dev Notes
Dev Notes: March 27, 2026
I've spent the last few weeks building startup hooks, context files, lock files, and keyword-matching scripts around Claude Code. The goal: every session starts warm. The agent knows what project I'm in, what happened last time, and which skills to suggest, before I type anything.
Then I read OpenAI's article on "Harness Engineering." GitHub published a similar one about multi-agent workflows failing. Both described exactly what I'd been building. A harness is the set of scripts, context files, and session patterns you wrap around an AI coding agent so it stops rediscovering your project from scratch every time. The hard problem isn't the models. It's everything around them: how agents get context, how sessions persist, how what one session learns transfers to the next.
Most developers have some version of this already. A CLAUDE.md file. A few rules. Maybe various uses of .claude, .agents, or .cursor directories. The difference between a scattered collection of config files and a harness is that a harness works as a system. Your startup hooks load context. Your lock files track which project is active. Your session notes carry state forward. This is the next layer of AI-assisted development, and most teams haven't started.
This Week on Slightly Caffeinated
E54: OpenClaw, Obsidian, and Dev Tools
TJ and I walked through the full OpenClaw + Obsidian setup: headless sync on a DigitalOcean droplet, Discord channels mapped to vault directories, and how the content flywheel runs day to day. We also got into Claude Code skill evals, API cost management (switching from Sonnet to Gemini Flash), and TJ's first two weeks at his new gig.
New episode drops tomorrow (Saturday) at 8:30am Eastern.
What I'm Learning
I built my harness in a few afternoons. Hooks, context files, lock files, a JSON mappings file. Works great for one person.
Then I tried to bring the same ideas to a team at work. Where do planning docs live? Who owns the startup scripts? What happens when someone builds tooling on their own time and the company wants it? Three engineers volunteered to pilot the approach. The first question wasn't technical. It was "do we have time allocated for this, or does it come out of feature work?"
The technical layer is ready. The organizational layer isn't.
Links Worth Your Time
Harness Engineering: Leveraging Codex in an Agent-First World - OpenAI's take on designing environments and feedback loops for coding agents. The shift from "make the model better" to "make the system around the model better" is where the industry is heading.
Multi-Agent Workflows Often Fail. Here's How to Engineer Ones That Don't. - GitHub's argument that multi-agent systems need explicit structure: typed schemas, constrained actions, and structured interfaces. Pairs well with the OpenAI piece.
I Built a Content Flywheel That Runs on Obsidian and AI Skills - New blog post this week. The Sunday night pipeline that syncs stats, generates reviews, and produces next week's content calendar.
Not Every Podcast Has a Transcript. Now I Don't Care. - Second blog post this week. URL in, structured Obsidian note out. Whisper handles the audio, Claude handles the structure.
Dev Tool of the Week
I built a two-hook system for my Obsidian vault that handles project context automatically.
A UserPromptSubmit hook reads every prompt I send and pattern-matches it against a JSON mappings file with 25+ project definitions. Each mapping has keywords, regex patterns, and a priority level. When I type something about "newsletter" or "devx," the hook figures out which project I mean, loads the right context files and last-session notes, suggests relevant skills, and writes a lock file so later prompts don't re-trigger detection.
A SessionStart hook checks for that lock file after /clear. If a project was explicitly loaded, it restores the full context stack automatically. Clear the session, keep the project.
The result: I switch between work projects, side projects, podcast prep, and newsletter drafting in the same vault without manually loading anything. The agent always knows where it is.
That's it for this week. If you're building your own AI development harness, or just starting to think about what that means, hit reply. I want to hear what patterns are working.
– Chris