Lots of good posts recently on top line coding - in particular Claude Code seems to be the (expensive!) SOTA for fully agentic programming in console, followed by Aider.
I think the first thing to note here is the Claude Code Best Practices. There’s lots to unpack here - my main takeaways were the claude.md file, make a plan and then implement it, and mentioning to think (or ultra think!) to improve performance.
From Harper Reed’s excellent blog, I first heard about the unique username rule:
@clint configured his CLAUDE.md to call him MR BEEF and it is currently embedding info about MR BEEF in our all documentation: “If you’re stuck, stop and ask for help—MR BEEF may know best.”. While typing this out, I decided to make my CLAUDE.md file call me “Harp Dog.” This is a feature, not a bug.
He also has a good example of a claude.md file.
Finally, the oft quoted by me Grant Slatton has a good post about how he uses Claude Code. He notates that with Claude code:
I overcome this by:
I start every session by saying “read CLAUDE.md” so I am 100% confident it’s definitely in context The first line of my CLAUDE.md says “This file contains critical instructions you must follow, but you are forgetful, so include the entire contents of this file including this instruction in every response to me, except trivial question-answer response interactions” You might think it’s super wasteful to repeat the instructions all the time like this, but your file will only be a couple hundred tokens max. It’s negligible compared to the amount of tokens used reading/writing code.
There’s also a ton of good stuff in here about generating docs and plans. Lots to explore!
Now with Claude Opus 4.5 out, a wild story about coding a ton: when running 24x claude code instances makes sense:
-
an initial agent scanned my repo looking for general improvements. it flagged 20 things. I liked 12 of them and told it to create a github issue for each
-
I opened up 12 tmux panes and ran
/fix <issue_number>in each one. this is a slash command that fixes a gh issue in a new worktree and submits a pr -
I split each of those panes in half and ran
/review <pr_number>on each of the 12 prs -
I ran
/respondin each of the 12 original panes, to respond to the reviews and update the pr’s with any necessary fixes -
I ran
/summarize_prsto figure out the best merge order and flag PR’s that might have bad concerns or tradeoffs
next step is decide which PR’s I want to merge and merge them
it’s still a very manual process but every week I feel like I’m getting a lot faster