I Deleted My Build Harness
The infrastructure I build around AI is obsolete every couple of months. The best thing I can do is float on top - know my own limits, and pin my process to Anthropic's best practice.
30 September 2026
I worked for weeks on a build harness that I ultimately deleted 95% of.
I have been hitting my stride with building recently. I've made over two thousand contributions in the past month across my work and personal GitHubs, and am consistently building across 6 or 7 work projects and 1 or 2 private projects. As I go through this volume of creation I've been careful to note where I am the limiting factor to my success and where the model is. The difference in my capacity to build ultimately now comes down to my ability to trust the outputs the models give me without extensive review (that I'm hardly qualified for anyway).
The infrastructure I've built to make creating with Claude Code more trustworthy and effective becomes obsolete every couple of months. I've finally learned my place within the process.
Initially I was the orchestrator, ferrying prompts between Claude Desktop and multiple Claude Code sessions - plan creation and validation, critics, revisers, cold adversarial reads, etc. - all shepherded by me. Then I built an orchestrator that would invoke those processes as subagents, running through all those steps, writing reports, handing off to the next agent, revising in loops. Then I discovered hooks and code tests and built a poor man's /code-review.
I have inefficiently chased efficiency, while admittedly getting significant gain along the way - in understanding but also execution.
What I've realised is that because AI advances so quickly the best thing I can do is float on top. Don't buy in too heavily into any particular process, and consistently audit my way of working to make sure that I'm on the leading edge of best practice.
I do that by first understanding what my own limitations are. What do I need from a workflow to make myself more effective and achieve the outcomes I'm looking for? I do a good job of defining what I wish existed, understand how things fit together, and have a good feel for process, order of operations, and my distance from desired outcome. I cannot smell-test lines of code for accuracy, suitability, or elegance.
Anthropic releases the needful so quickly that I've realised that, except where my process is extremely nuanced, it's best for me to solely focus on pinning my process to their best practice. I've got the build primitive finder that sources large-scale patterns, and now I've got a cron job that audits my skills against best practice.
I've come out of it with a build harness that I am happy to use (at least for the next couple of months), which I'm sharing here.