I keep forgetting to add the changes I make on the website in the /changelog of the website. The changelog is a great way to keep track of the evolution of my website and I want to keep tracking changes on a weekly basis.
The speed at which Anthropic is releasing features for Claude is fast. My guess is that they already have an internal Opus 5.0 model for coding running.
The release and subsequent adoption of OpenClaw was a revolution in the agentic AI world. Anthropic is currently doing all they can to try to bridge the gap to have Claude be able to be an agent with similar value as OpenClaw. OpenAI reeling in Peter Steinberger (and OpenClaw) must have accelerated these efforts.
In today’s experiment I am going to try out the /schedule feature of Claude Code. I am going to compare its ease of setup and how much intervention it needs from me.
I started by scheduling a Claude Code daily task to go through the meaningful changes on this website and drop them in the changelog.
Claude Code then started going through the task as it would normally do and did some reasoning about things like what commits to include and how to write the changelog based on previous entries.
The biggest difference with OpenClaw and its cron jobs became clear. It suggested an initial manual run to make sure all security gates were lifted. There were quite a few ‘allows’ needed to give permission to read/write the changelog and push the changes automatically.
In OpenClaw you just ask to do it, and it works automatically. All without any intervention. OpenClaw has tightened security recently and still doesn’t require a manual first run.
The task appeared in the overview that Claude created and I was able to manually run it there. A well designed screen to handle this.
From now on I have a daily task that is scheduled at 10PM to just go through the changes (if any) and adds them in the changelog automatically. As I keep forgetting to add them manually.
However when it ran automatically the first time it did run into a problem, as Claude Code still had some missing permissions.
Tool permission request failed: Error: Tool permission stream closed before response received
I thought I had allowed everything needed, so I had to fix this. Quite annoying if you are used to OpenClaw, as it keeps manual intervention to the bare minimum.
Key Insight:
Striking a balance between ease of use and security must be a very tight rope to walk. I get why Anthropic and Claude have to do this, but the execution is still broken. If the whole point of the manual run is to catch permission issues upfront, and it still fails on the first real run, then the way of handling security is not executed well. The extra friction didn’t make it more reliable.