My AI coding setup has been evolving the last few months. Here are some things I’ve tried and moved on from:

  • Gas Town. It’s just too much to wrap my head around right now.
  • Pi and custom harnesses. I’m not at the stage where I need to deeply customize my workflow, starting from zero. I’m also not at the stage where I’m frequently switching models for different tasks. The Claude CLI is still good for me at this stage.

Here’s my current simple-to-use setup today:

  • Claude Code CLI. I use a variety of plugins with Claude Code. I don’t use IDE integrations or the Claude desktop app.
    • superpowers provides a basic software development framework that really improves the quality of initial implementation. I don’t even think about this one, it gets invoked automatically all the time.
    • autoresearch provides a simple optimization loop for measurable tasks. I recently used it to cut down bundle sizes in one of the programs my team ships.
    • Slack’s official plugin for reading messages.
    • gdocs-cli. Anything at work that doesn’t happen in Slack happens in Google Drive.
    • Grafana-internal skills and tools for our wiki and queries.
    • bugbot-loop listens for PR comments and responds to them by implementing good suggestions and resolving bad ones, over and over.
  • Cursor BugBot installed in our Github repositories. This is the tool that I think provides the right combination of ease and high quality in its review processes. Combining this with bugbot-loop immediately increases PR quality without intervention on the part of any humans.
  • agent-deck. A Tmux layer that provides management of multiple agent sessions. It understands git, especially worktrees, and watches the state of sessions. This way I can tell which of my sessions is still churning, and which are idly waiting for intput.
  • ars contexta. This one is a little complicated to explain, but it’s a knowledge base that an agent automatically maintains and refines over time.

So, my work setup tends to look like this: I have a custom skill for summarizing Slack that runs once a day via launchctl. It DMs me a summary each morning that helps me build up my to-do list. Throughout the day I throw my notes into ars contexta. I used to use Logseq for this, but I’m doing less manual creation of notes now.

I have an organized set of sessions in agent-deck.

  • A section of coding tasks - each associated with a set of PRs. When the PRs are finished to main they’re killed and cleaned up. After the initial coding, when the PR is updated I start /bugbot-loop for up to 10 iterations.
  • A set of longer-lived “research” sessions where I’m keeping notes about and getting feedback on larger ideas and direction for projects. These usually end up writing markdown files that are consumed by ars contexta or written into things like Architecture Decision Records for a given project.
  • a single ars contexta session where I throw in notes. I can also ask this to search recent Slack activity or gdocs for information to add to notes.

As far as concepts and tools I’m interested in investigating to continue improving my workflow:

  • agent-deck “conductors” to increase autonomy for task-oriented sessions even more.
  • Archon. Providing a little more determinism in my software workflows could be nice. Compare “fix a bug” to “explore the codebase for improvements”. These practices should probably look different, and . Seems like a promising way to increase autonomy with some of these tasks.
  • Spec-driven development. I need to dedicate more time to reading and prototyping here. I’m optimistic about improving autonomy in my sessions via more detailed planning.
  • Context graphs. The concept feels like it may be able to encode a lot of “Chesterton’s Fence” information that agents struggle with. In particular I’m interested in it from an Observability standpoint. Every running software system is unique, and we’ve never been able to establish industry-wide practices that everyone follows to the letter. Semantic Conventions and the like are moving Observability in that direction, but discovering structure is somewhere that LLMs are very promising for.