ChatGPT Briefings Via RSS

I’ve had some success getting daily briefings from ChatGPT in the form of an RSS feed. What I did was stood up a Hugo site, backed by Git, and wrote an AGENTS.md file with instructions on how to use it. Basically, checkout main, create a new post, commit, then push to main. Pushing would run the CI/CD pipeline and publish the site.

Really nothing fancy, although I did write a couple of tiny Go apps for the agents to use. One is for creating new posts, and all it does is ensures the directory for the post is present and that the filename and front-matter is how I like it. There’s also a tool the agent is instructed to run to “finish” the post, which does things like replace Twitter links to XCancel. It was easier doing these deterministic operations in code, rather than come up with suitable instructions for the agent. How to use them is documented in AGENTS.md, and it seems like the agents are using them correctly.

Then, modelling after Manton’s prompt, I setup a schedule for Codex to get AI news from Twitter and write it to the Hugo site:

Find the latest public posts/tweets from notable OpenAI and Anthropic employees, summarize the most important updates, and include links to the original posts. Prioritize posts from the last 8 hours and focus on AI product, research, developer, policy, and company-relevant updates.

Save your responses to the repository at @dir/to/hugo-site as per the instructions in AGENTS.md of that project. Use the category ai-updates.

You have full permission to commit and push to this Git repository.

I let it run overnight, and the next morning, I got an update in Inkwell. This continued for a few more nights, until a scheduled OS restart caused the whole thing to hang. But it’s back as of this morning.

One surprising thing I found the agent doing is relying on the previous runs to know what it should include in the latest one. I can’t tell if it’s getting this from it’s memory or from the Hugo markdown files. I’m guessing if I asked it to consider the prior posts in its decision of what to include in this run, it will do so. A happy accident from relying on Hugo for this.

I’m slowly trying to extend this to other areas of interest. Tweets from AI employees might be nice, but my ultimate aim is to reproduce something like ChatGPT Pulse. The open question remains whether it could use my chat history. The fact that Codex could do this was never in doubt: the issue was agents lacking access to my previous chats. OpenAI’s separation of products continue to baffle me, but it seems like my ChatGPT history is now available to ChatGPT “Work”. So I re-enabled and adjusted the scheduled I created when I previously attempted this and gave it a test. It ran successfully, and produced a briefing, yet it didn’t seem to have picked up anything from my history.

Anyway, we’ll see how it goes.