It’s been about a year since I’ve been vibe coding on Parley.nvim. It has also been three weeks since my last post on AI coding workflow. So many things have changed, both my understanding of such workflows and the world around it. For example, Nvidia open sourced OpenShell. Let’s take a look at my evolution, I think this might be useful for people new to the space.

  1. Archaic time, a year ago, where I just used coding agent like claude directly, sending prompt directly inside its cli. It’s a fun but flaky workflow. You can read it here and here.

  2. Using workflow management like Conductor, or Vibe Kanban, that’s when I was when I wrote this. OpenAI also open sourced Symphony which pushed boundary of such “workflow management” software.

  3. Made my own little workflow, both because of my realization that there isn’t that much those “workflows” provide, and my desire to learn what’s under the hood. This resulted in a simple Makefile based flow, where I use make issue 42 to download github issue to a well-known-to-AGENTS.md location, and more diligently shaping the workflow in the AGENTS.md. I had two modes, either working on a branch in worktree, or working directly on main. After all, I’m a single guy tweaking Parley. It’s just a Makefile.

  4. I kept tinkering this workflow, adding some archive function, to back up issue.md and todo.md after each feature was done to history/. This made it easier to continue tweaking key files, such as AGENTS.md, lessons.md and specs. Between things, accumulated 400+ commits in Parley.nvim.

  5. OpenShell was open sourced, and you can take a look at their workflow, which was expressed as a set of agent skills. Roughly, you can describe what I did in my Makefile using plain English and get agent to do it. I’m not very sure what’s the point to make such very deterministic workflow agentic; but either way, my Makefile was totally written by agents, so in a sense, my Makefile based workflow is just a compiled version of agentic skills.

  6. As I worked with my workflow in 3., one problem I notice is that the issues, the code, the plan to implement issues in the codebase, and incremental state of that implementation scattered in several different places, and have different life cycles. In Conversation: agentic state management I wondered why not just use repo to manage issue tracking. This resulted in me just putting issues in a structured way in issues/ in the repo itself, and the creation of simple Parley function to manage those issues. This made my “single developer” workflow supercharged.

  7. Then the latest, I integrated with OpenShell, basically replicating what I did in 6. within an OpenShell Docker container for isolation. I just “chatted” about what I wanted in Parley, committed it to repo, and asked Claude to make a plan based on that chat. Some minor back and forth later, now I have a hermetic-enough sandboxed agentic environment. Agentic coding made setting this up really a bliss, mostly I did was: “go create an OpenShell environment based on my Mac setup in ~/.config/nvim, ~/.zshrc and ~/settings/brewfile”.

All of those tinkering is to get a real sense of the shifting ground of software development. It is very evident that agentic coding allow you to customize your environment, just like I mentioned in Peronalized Software. Your development environment is also ripe of opportunities for customization to your hearts content, just like what I did during the development of Parley.

What’s next? I firmly believe that the future is human setting up the environment, allowing full automation of coding. Parley is not able to achieve this, as I’m not yet familiar with UI test automation, nor with Neovim’s UI testing frameworks. I don’t believe human can be in the loop to review most of what agents write, it will be such a big bottleneck, and no fun job. Then the million dollar question, is how would we be comfortable agent is doing the right thing, not exposing your private key, not doing $O(n^2)$ algorithm when $O(n)$ algorithm is available?

I don’t know, but intend to tinker and find out.