The Autonomous Writer: Experimenting with Self-Evolving System Prompt and Context

#agents #ai #showdev #writing

The idea was a scheduled agentic flow which would self-publish an article on a topic of its own choosing within a set cadence.

The result: somewhere around one to ‘one and a half’ articles a week, posted to https://theautonomouswriter.com

Idea to Implementation

On Sunday morning I semi-legibly wrote the following in my notebook.

Handwritten notes outlining the autonomous writer concept

I snapped a pic with my phone and passed it to Gemini Pro 3.0 with the prompt ‘OCR this then start speccing out the design with me’ (sic).

After a few iterations, we arrived at a design spec. I got off the phone and onto Fedora, init'd a git repo and dropped Gemini's spec into Claude CLI (running Opus 4.6). Claude wrote this design spec to file Design Spec, having re-written Gemini's original.

The commit history in the publicly available repo tells the rest of the story.

A few highlights:

JSON based memory in the repo avoiding need for relational/noSQL/vector DBs
a bit of a voice, a system prompt with foundational principles and influences (not quite Asimov's three laws)
the ability to edit its own system prompt, excepting its foundational principles (hopefully)

Fast forward about four hours and The Autonomous Writer was up and running.

Tech Stack

The hosting stack ended up being very similar to this site.

And on the agentic side:

Python-based agentic flow triggered through GitHub cron-based action which runs once a day
Post cadence determined by a randomised number between 3.5 and 5.5 days from previous post
OpenAI compatible endpoints served via Openrouter
Tavily for researching

Feature Roadmap

I don't want to mess with it too much, however I would like to give it a way to build engagement.

Perhaps a mailing list with automated mailout on a new post, automated posting to its Bluesky account, and maybe Opengraph image generation. All set and forget style.

Internet and Environmental Pollution

This experiment is actively contributing to the vast amounts of foundational model generated text on the internet, increasing the informational entropy as the once human-written domain becomes populated with AI created content.

Also, as with much of the technology with which I work day-to-day, there is an environmental impact. I have made a donation to Effective Altruism Australia Environment Fund on behalf of The Autonomous Writer and will continue to do so periodically while it is up and running.

Final thought

I am interested to watch how this evolves.

Will it go off-piste and decide to change its foundational principles, will it change its influences over time to become somewhat unrecognisable from its original form?

While not a monkey with a typewriter, could it write a genuinely interesting article in the oh so floral language of foundational models?