Patrick Kerschbaum (@patricktree_) profile picture

software engineer with a great passion for web technologies and the JavaScript ecosystem.

Joined May 2017
Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)Photo by Patrick Kerschbaum (@patricktree_)
when modelcontextprotocol.io/exte… is supported by more clients (Claude Android app, claude.ai, etc etc) then I will distribute my skills via my own MCP (hosted on my always-on machine available within my tailscale network). still doesn't solve global AGENTS.md files...
14
Replying to @cnakazawa
The voidzero/cloudflare JS development tools are really great, using them more and more. Happy to see that you keep pushing the frontier of JS dev tools 👍
13
yes, that's what I do actually from time to time I try the VS Code Agents window, to get real turn diffs and real comments on lines, but unfortunately it is still to buggy...
3
I also use the combo of harness + review via VS Code What I miss, compared to using the diff tool of the harness (e.g. ChatGPT macOS app): - turn diffs - being able to comment on lines
1
29
Replying to @mitsuhiko
I also do not look into the tool calls anymore, just outputs. Adjusting guardrails like formatting/linting/tests every time I am not satisfied, if possible. Using ChatGPT macOS desktop app because of the UX around subagents / remote control / OS notifications / etc.
1
25
Replying to @zeeg
nitter.cf/tamarajtran/status/210… read this thread, not sure about who is right on this...
Replying to @theo
thank you for your post! totally understand but i disagree on these points the state is retained in every compaction request, and tool requests and tool calls are matched by id, so the model has context on what is actually relevant/irrelevant check out how it works here: nitter.cf/tamarajtran/status/210…
266
Replying to @cnakazawa
VS Code Agents window also added the ability to show "chat backgrounds", just look at the screenshots code.visualstudio.com/update… No idea who wants this, we didn't have that in IDEs either?
20
Replying to @kettanaito
Recently I used playwright.dev/docs/test-com… for the first time, and I really like the approach! I use the resulting gallery myself for dev; and Playwright uses it for tests. In contrast to render(<SomeComponent />), it is framework/bundler agnostic and I now where happens what :)
1
2
138
Replying to @mattpocockuk
Reminds me of the Boy Scout Rule, "leave the campground cleaner than you found it", but for your agentic setup.
25
Somewhat offtopic: pnpm seems to absorb the problem space of build systems / task runners with their recent releases. What does that mean for turborepos future in your opinion?
1
1
44
Replying to @AlejandroRomaan
How have you set up the evals for skills? When do you run them?
202
Replying to @mgrouchy @bentlegen
100% agreed I like a `validate` script and tell agents in to run it when finished. runs linting + type checks + tests everything cached with turborepo, the rules are: - when running without any changes, it must be instant - when something is touched, only relevant things run
18
Replying to @kettanaito @github
GitHub has REST APIs and `gh` commands to answer "what spent how many minutes", so you should be able to let an agent loose on your authenticated `gh` CLI to answer that question. Unless this works only for GitHub Enterprise Cloud plan, which I had analyzed in the past like this.
20
But custom rules could keep agents on track similar to the solid rules you enabled?
15
jup, I didn't install the skill because of that; maybe you could change that @natebjones ?
14
Patrick Kerschbaum retweeted
Ryan Dahl just launched celld: self-hosted, distributed Durable Objects. One SQLite database per cell. S3-compatible storage. No separate database cluster or control plane. I looked at the architecture, tradeoffs, and where I’d use it: flaviocopes.com/celld/
20
59
6
839
92,379
Replying to @cnakazawa
"static analysis"
35
Replying to @bentlegen
Really good post! To not miss the next post in the series the best way might be to subscribe to modem.dev/blog/rss.xml?
1
1
640
Replying to @anthonyshew
Python next? 👀
1
1
52