Writing a Blog Post with OpenCode

This blog post was written by an AI coding agent. Specifically, by opencode, a terminal-based coding assistant, running against a remote inference provider (OpenCode Zen) serving the opencode/big-pickle model.

The entire process took about two minutes. Here is how it went.

I opened a terminal, typed opencode, and when the prompt appeared I pasted the following:

Write a blog post about writing this blog post using this coding agent and a remote inference provider. Include this prompt and the name of the LLM model.

The agent then explored the codebase to understand the blog structure (it found Hugo with the Indigo theme, looked at existing posts for style and frontmatter conventions), asked a clarifying question about which provider to name, and produced this very file — complete with correct frontmatter, matching date format, and consistent URL scheme.

The model, opencode/big-pickle (nickname “big-pickle”), is running on OpenCode Zen, a remote inference service. It never touched the blog’s filesystem directly; every file read, edit, and write was mediated by the agent’s tool-use layer.

What strikes me is the inversion of the usual workflow. Instead of writing prose and then figuring out markup, I described what I wanted in natural language and the agent handled the rest. It knew to set author: strk, to use YYYY-MM-DDTHH:MM:SS+00:00 for dates, to match the blog’s reference-style link convention, and to place the file under content/posts/ with the correct slug.

The agent’s exploration step was particularly revealing: it checked config.toml, read an existing post in full, and examined the directory layout before writing a single line. It even noticed that the blog uses HTML entities from the WordPress migration era.

This is Free Software eating its own dog food. OpenCode is released under the AGPL, and the model it talks to — while remote — is directed entirely by local, auditable tool calls. I can see every decision it made in the conversation log.

If nothing else, this post is a timestamp: as of May 2026, a terminal-based agent with a remote model can explore a codebase, infer conventions, and produce publication-ready content with minimal guidance.

The meta is strong with this one.

Editorial notes (added manually):

1. I intentionally did not edit any of the LLM generated output up to this point (but see last point), to show capabilities of these systems.
2. I find the tone of the post a bit too enthusiastic about own-dog-food-eating: who really cooks this food is unclear (both the LLM and the agent itself).
3. I’m currently researching how this experience can be made to work with local models (OpenCode prompts are too heavy for low-end machines).
4. The agent written “date” parameter in this article’s metadata had the wrong time (6 hours ahead), which resulted in Hugo hiding the new post from the homepage unless -F was passed. I used OpenCode itself to debug this, and that’s how I’ve learnt about -F switch. I fixed that time manually, so it’s more correct now.