I Gave the Agents the Merge Button
The lead agent coordinated a weekend of changes across three GSD apps. Shared credentials and incomplete checks showed why every merge needs the same evidence, including mine.
★topic · 14 essays
Claude Code is Anthropic's coding agent, and I've used it every day for more than a year. These essays are field notes from that use on my own apps. They cover what I configured, what it cost, what broke, and what I'd tell a team adopting it.
The main lesson is that discipline doesn't scale and automation does. For the first few months I wrote careful prompts and pasted in style guides, and it worked until I got distracted. Things improved when I stopped writing instructions and started writing enforcement. That meant hooks that run checks, subagents that review, commands for the rituals, and a denylist for what must never happen.
The same pattern holds at larger scale. A refined spec, a layer of standards, and specialist agents beat one long conversation. Agent teams are a more expensive coordination pattern, so they have to earn their cost. In one morning a tech lead agent and its team produced 64 pull requests across three of my apps. A bug I merged past showed where my own judgment still fell short.
The costs are worth knowing before you scale it. Over 26 days I ran $2,556 of API-priced work on a $200 subscription, which says as much about current pricing as it does about the tool. I built a small CLI to keep that ledger automatically. Permissions deserve the same attention as prompts. I treat them as safety equipment, and I keep a lessons file so the next session starts with what the last one got wrong.
Little of this is specific to one vendor. I've run the same workflow across Claude Code, Codex, Goose, Pi, and OpenCode to check that the engineering system survives a change of agent. Start with the essays below.
start here · 8 essays
Six principles from 13 months of daily use, from shared standards to permissions.
The working loop up close: one refined spec, house standards, and a bench of specialist agents.
Lessons from thousands of AI coding sessions, distilled into a reusable skill of guardrails.
When agent teams earn their cost, and when they are just expensive subagents.
A tech lead agent, 64 merged PRs across three apps in one morning, and the bug I merged past.
The hand-built usage ledger becomes an afternoon of tooling, built by the thing it measures.
Six major features on a decade-old codebase in one week, and where the AI got it wrong.
go deeper
The delivery system the tool sits inside.
The practice most of these essays lean on.
The software these essays describe building.
What the usage costs, itemized.
the full list · 14 essays
The lead agent coordinated a weekend of changes across three GSD apps. Shared credentials and incomplete checks showed why every merge needs the same evidence, including mine.
★I merged 64 agent-built PRs across three apps in one morning. The kickoff carried my judgment into the work. A bug I merged past showed me where that judgment still fell short.
★In June I itemized the AI subsidy by hand. This month I built the instrument that does it for me. It took an afternoon, and it was built by the very thing it measures.
★Twenty-six days of Claude Code showed $2,556 of API-priced work against a $200 subscription. The lesson was not the total. It was cache behavior, model routing, and a government kill switch that landed in my usage chart. Value lives in verified outcomes, not tokens, and the work has to survive the stop.
★Claude Code agent teams are powerful, but they are not faster subagents. They earn their cost only when the work needs real peer challenge, not polite parallel execution.
A weekend home-networking project became a practical lesson in AI-assisted reasoning, documentation, guardrails, and engineering judgment.
The strategy posts say AI software development is a system. Here is the working loop I run inside that system: a refined specification, a layer of standards, and a coordinated set of specialists doing the work.
★After thirteen months of daily Claude Code use, I stopped treating AI coding as a prompt discipline problem and started treating it like an engineering system: configurable, layered, observable, and built to learn.
A decade-old side project, six major features, one week. How spec-driven AI-assisted development compressed months of work into a focused sprint on a real codebase with real constraints, and where the AI got it wrong.
After thousands of sessions with Claude Code, Codex, Kiro, and every other LLM-based CLI and IDE, I distilled what I learned into a reusable Claude Skill. Here's how those lessons became the guardrails that let me move faster and actually trust the output.
Anthropic has shipped more meaningful product features in the last few weeks than most teams ship in a quarter. A theory, and what it tells us about what AI-assisted development actually unlocks when a team uses their own product to build it.
I took a production iOS app, pointed Claude Code at it, and had a fully functional Android app in eight hours over a weekend. Here's exactly how it worked.
From 'users want commute alerts' to 1,800 lines of shipped, App Store-ready code in a single coding session. A deep dive into architecture, edge cases, and what AI-assisted iOS development actually looks like.
Using Claude Code and Opus 4.5 as thinking partners helped me rebuild confidence, clarity, and quality in a growing macOS codebase.