Happy Friday friends!
My Claude limits reset today so I am looking forward to having the privilege of using Fable again for a day.
It has been a while since the last newsletter. For those of you who are new or have forgotten who I am - let me reintroduce myself.
I am Abhishek - I drank the Claude kool-aid last year and have been writing about my experiences using Claude Code.
My laptop became a bottleneck
How I write code has changed in the last 6-9 months. A coding task can run from 20 minutes to a couple of hours.
I would only run my longer coding tasks when I knew my laptop would stay open. I used caffeinate on my Mac to prevent it from going to sleep. I would keep my laptop lid half open when commuting. This went on for a couple of months.
Then came the breaking point - memory & disk errors. Old sessions and worktrees kept piling up.
I came across Orca IDE around this time on X. What got my attention:
It runs Claude Code / Codex in a terminal so the interface is a lot more familiar (unlike Conductor)
Support for running Orca both locally and on a remote server
Native iOS app
I already had a Hetzner VPS from when I was testing OpenClaw. Setup was easy: I pointed Claude at Orca's docs, had it install the Orca server on the Hetzner box, and added that server to my local Orca config.
I eventually added all the configs to Terraform because I ended up getting a larger VPS and didn’t want to go through the manual process all over again.
Later I moved it all to Terraform. I upgraded to a bigger VPS and didn't want to redo the setup by hand.
Orca’s local + remote support helped me move over my workload gradually.
Whats working well
All of my coding work now runs on Hetzner.
A task keeps running even if my laptop is closed. And I can review work from my phone using the Orca app if I don’t have access to my laptop (my son’s swim lessons).
Anthropic explained how Claude Code tokens really work
Anthropic published a post last week on getting more out of your Claude Code sessions. It explains what makes a session expensive, and what to do about it. It's one of the most useful things they've written.
First, the pricing basics from the post:
Output tokens (what Claude writes) cost about 5x more than input tokens (what Claude reads).
When part of your session is cached, reading it again costs only 0.1x the normal price.
Writing to the cache costs up to 2x. You pay that once, then every turn after it is cheap.
Here are the key takeaways:
Run /clear when you start a new task. By turn 40, Claude re-reads the 39 turns before it. Old context makes every new message cost more. Run /rename first if you may want the session back later.
Pick your model and effort level at the start. /model, /effort, and fast mode are part of the cache. Change one mid-session and you pay full price to rebuild it.
Use /rewind instead of /compact to undo recent turns. Rewind keeps the cache. Compact throws it away.
@-mention files instead of typing the path. This skips a Read call. You only need to do it once — the file stays in the session.
Add quiet flags to noisy commands. Put your daily commands in CLAUDE.md with the flags included. Command output under 30,000 characters stays in the session until the end.
Run /compact before a long break. The cache expires after 1 hour on subscriptions (5 minutes on API keys). Compacting is much cheaper while the cache is still warm.
The memes
Opus 5 communicating in Claudish
That’s the week.
Abhishek