May 5, 2025From Agents to IntentMy coding agent could scaffold applications and pass CI. Then I asked it to modify something, and everything broke. That failure led me toward a different source of truth: structured intent that AI can read, rather than infer from implementation.
May 5, 2025Generic BuilderAn intent-based toolkit needs specialized builders—but what builds the builders? The Generic Builder proposes a configurable starting point that can create missing tools, with a router agent connecting conversational requests to structured intent files.
May 1, 2025Sufficient SpecificationsA specification can become as complex as the code it replaces. “Sufficient Specification” proposes a different balance: let LLMs infer familiar details, then add explicit requirements where the generated implementation falls short.
May 1, 2025Recursive IntentFeeding generated specifications back into a toolkit could grow a one-line CRM brief into models, APIs, and pages. But each pass can amplify mistakes too; recursive intent needs ways to preserve decisions and prune unwanted growth.
March 28, 2025Crafting READMEs for AIEvery time an AI agent has to reverse-engineer your library, you pay for the lesson in tokens. A clear, versioned README lets you write that knowledge once—especially for code the model has never seen.
March 25, 2025The Rise of "Test Theater"AI coding assistants are excellent at generating tests with high coverage. But there is often a critical problem: they are testing that code behaves as written, not as intended. This "Test Theater" creates a dangerous illusion of quality while providing almost no real protection against bugs.
March 21, 2025Minimal Agentic CoderA coding agent can fit in around 200 lines of TypeScript: ask Claude for shell commands, execute them, and feed back the results. What does that tiny loop accomplish, and where does it fall short?
March 20, 2025Building an Agentic CoderI gave an LLM a shell and a few file tools, and it started writing and debugging its own code. Making that prototype useful meant solving less glamorous problems: blocked commands, oversized files, and a context window that kept filling up.
March 17, 2025Software as EphemeralIf rebuilding becomes cheaper than maintaining, the code may stop being the asset worth preserving. Requirements and behavioral tests could become the durable core, letting us choose tools for today and replace the implementation later.
March 13, 2025Automating Agentic CodersBefore taking my daughter to daycare, I turned two GitHub issues into feature PRs from my phone, without writing code. A few issue comments were enough for MyCoder to plan and implement the changes through GitHub Actions.
March 7, 2025GitHub Mode for Agentic CodingGiving MyCoder GitHub issues, pull requests, and CI logs changed what I needed to supervise. The surprising part was GitHub’s role as persistent memory: I could leave the agent working for 30 minutes and review its PRs asynchronously.
March 6, 2025Agentic Mistakes Improve CodeAdapting your codebase to match AI coder assumptions doesn't just help AI -- it uncovers hidden complexity, ultimately benefiting human developers.
