Dylan Watt

Using Devcontainers to Fix Coding Agent's Foibles

Wed, 20 Aug 2025 00:00:00 -0700

I’m publishing the template I use for AI Coded projects here.

For the last two years, I’ve been forcing myself to use AI to write as much of my code as possible, because I want to understand the technology’s weaknesses and strengths. Things started to take off with Cursor’s agent mode, and the release of Anthropics Sonnet 3.7 model and then Claude Code rocketed it to a place where it could write all of my code… with caveats.

5 Essentials for getting the most from Coding Agents

Thu, 12 Jun 2025 00:00:00 -0700

5 Essentials for getting the most from Coding Agents

Spend the time to teach them your Project

There are a lot of AI coding agents out there, but they all stall out if you don’t tool your codebase for them. There’s a ceiling of how effective they can be if they’re running in a boilerplate environment, with no customized context or guidance.

MCP's Achilles Heel

Tue, 03 Jun 2025 00:00:00 -0700

I’ve been getting into personalizing Claude desktop for more much more detailed, powerful responses.

Model Contect Protocol (MCP), is a way to expose tools (APIs) to an AI agent to get custom data, or do custom actions. The issue is that the request and response are all fully written and read by the LLM, so for large datasets, you’re making the LLM your data transport layer as well as an AI.

Letting AIs run my Computer

Wed, 21 May 2025 04:00:00 -0700

I recently had a well upvoted comment on hacker news regarding an AI agent and, as is the way of the internet, somebody decided to stalk my online presence to find something unrelated to be mean about.

PR for a Ticket

Tue, 18 Mar 2025 00:00:00 -0700

The City upon a hill for AI coding is that a well specified ticket should get an automated pull request implementing it. The act of coding itself will go from carpentry to CNC, where you still must understand the properties of the material you work with, but shaping that material is a easily repeatable task, so you just focus on the goal.

Many of the day to day tickets a software engineer runs into are already at the point where a well instrumented AI Agent can handle the changes. Updating a title, creating a well specified screen, and even implementing a simple CRUD endpoint are almost fully solved. However, many tasks, like creating a database migration, managing security concerns, and other tricky tasks are beyond current agents.

AI Rekindled my Joy of Coding

Thu, 13 Feb 2025 00:00:00 -0700

When I first started writing code, I had a sense of joy in the fact I was learning the arcane incantations that brought modern day magic to life. Over time that feeling faded as I faced the realities of building code at scale. Bogged down by so many opinions of the “right” way to do things, instead of the joy of creation I now felt the dread of knowing I hadn’t tested the code well enough, or that dependencies weren’t injected, or that I had 327 npm deprecation warnings.

Mon, 01 Jan 0001 00:00:00 +0000

About Me

I’m a software technologist with a focus on end user product. I’ve worked in VC startups for around 15 years, including my own. I’ve managed multiple teams, working across Education, Real Estate, Marketing, and Account Management, both B2B and DTC.

I’m currently working as a Fractional CTO/CDO for a few startups and non profits, heavily focused on the “Modern Data Stack” space, and for personal focus, I’m heavily invested in figuring out how to leverage AI tools to speed up development, both for my own work, and how to build Agentic systems.

Mon, 01 Jan 0001 00:00:00 +0000

The 8 factor AI Coding Guide

Test: I must have a fast, comprehensive suite of tests. Fast feedback loops are paramount. The longer an Agent continues down a path with a bad assumption, the harder it is to recover.
Modularize: I must architect my code modularly, as to minimize the context needed by the Agent.
Sandbox: I must run code in a sandbox that allows the agent to run wild. If I need to approve every tool use, I’ll never unlock its full potential.
Manage Processes: My code should run in a separate thread from my agent. IE if I work on a web server, the agent can’t start it as that will hang the agent. Live reload it in another thread, and send the logs somewhere the agent can check.
Automate: I must make my development environment trivial to spin up, so that I can have many copies going simultaneously for parallel AI development.
Document: Planning is more important than coding. Every code change should be preceded by the generation of a clear step by step plan with outcomes the agent can follow. That plan will persist across sessions, and act as long term memory for the agent. Ideally use Claude Code, Gemini, and Codex to check each others work when planning
Tool: MCPs supercharge my agent by allowing it to view my project with playwright, load additional information with websearches, and generally have more deterministic tools to take action.
Contract: I must use a statically typed language that can be quickly compiled to find errors. Interfaces limit the surface area needed to be loaded into context, and compilations find a class of bugs AI can create.