Process update - How I use Agents as of Dec. 2025

First, I want to start by mentioning that I failed in my initial goal to post at least once a month. The more ambitious version of me wanted to do every week, but alas I fell out of the habit as work/life demanded more of my time.

Plus, I haven't had a revelation in process since the early fall. New models have come and gone, and the hype cycles are growing shorter. GPT 5.1 was quickly overshadowed by Gemini 3 and then rapidly followed by Opus 4.5. I watched on with curiosity, read all of the comments on HackerNews, and then went back to shipping code using my trusted workflow.

Since my last post in September, I have heavily modified my approach towards planning and task breakdown. Initially, I would create a GitHub Project and then the issues related to that Project. It worked, but I didn't really enjoy spending time in the GitHub interface myself. I am a solo developer on my personal projects and everything is committed directly to main. My goal is to move fast - managing pull requests and remote artifacts reduces my velocity.

So I did the obvious thing and built a small CLI tool for managing a local-only Kanban board. It's super simple:

• Your GitHub root contains a directory for task and epic management (.tasks/)
• There is a subdirectory for each status and a special directory for epics
• Within each directory are markdown files with an outline of the task
• The content of the task is arbitrary; only the frontmatter is required
• The tool abstracts the management of the tasks, but ultimately it's all just manipulation of files

My process remains the same as before with a heavy emphasis on planning out the broader vision first (the epic) and then breaking it down into individual units of work (the task). The ultimate goal of any given task is for it to contain enough context, and links to additional context-bearing artifacts, that an agent has everything it needs to begin work.

Once planning is complete, my daily routine is to move a handful of tasks out of the backlog. Then, I use my custom feature-dev slash command to trigger Claude to work on a task.

What about Gemini 3? Or GPT 5.1? Or all the enhancement to Cursor? Have they not changed your workflow?

Honestly, no. I still much prefer CLI tools (so no Cursor) and the benchmarks seem to show that we are hitting a plateau of performance. I honestly have not felt the need to chase after the latest models as my workflow can produce the level of consistency I need to deliver software. Every demo I see for a new model involves people "one-shotting" various tasks which is not how I approach work.

However, I am seriously exploring moving full-time to OpenCode. The Claude Code interface is honestly a mess. Plus, I could take advantage of different models instead of relying purely on Anthropic models. My Christmas gift to myself will be to migrate all of my process commands over to OpenCode so I can start the new year with a fresh system for work.

Hopefully it will be one that isn't superseded within a few months!