Lukas Trumm

Lukas Trumm

@lukastrumm.com

Husband | Father | Software engineer | Scout leader | Endurance runner | Craftsman

26 followers65 following50 posts

Tip no. 6, use agents to filter what reaches you. Mine builds a daily feed from blogs and social posts that I read on an e-ink tablet, no ads, no irrelevant posts. And a skill reads my config and picks out the Claude Code changes relevant to me: github.com/lttr/claude-...

Tip no. 3, save 1/3 of the initial context in every Claude session: each tool and feature costs tokens at startup. Turn off what you rarely use in settings.json and stay in the smart zone longer. Any feature can come back per session. My setup: github.com/lttr/dotfile...

We assemble context for AI, but our brain needs onboarding too. The reverse context management move: instead of preparing context for the AI, have the AI prepare context for you.

Create a custom command or skill that tells you what you were working on, the relevant state, and how you'd likely want to continue. What matters is getting back into unfinished work quickly and without friction.

Claude Code is just an automated copy paste machine. But considering it can copy from almost anything and paste almost anywhere - its quite good 😀

Today I learned: AX stands for Agent eXperience. Well it actually make some sense. As a developer I want my AI agent to have great experience using my tools on my behalf...

Counting rows on an unusual chessboard seems simple since you sense there's a trick. But I couldn't get any major LLM to answer correctly without big hints. Do you have similar examples?

Lukas Trumm reposted

One of the reasons I like using Gemini 2.5 Pro for tech specs is that it can defend its decisions. If I ask, "why didn't you X?" it will give me the rationale and let me debate back and forth. Compare that to Claude: "You're absolutely right! I should have done X." Much less helpful.

What is all the rage about parallel AI agents? And how is it great to let the agent work for hours? I get that it solves the current problems, but overall it feels wierd. Why? Because humans can't really think in parallel. 1/2

One focus on one problem at a time and deal with tasks sequentially. When one wants to check the output of an agent and interact with it, parallelism is not helpful. What is helpful is to get the results fast, so you can focus on the next task and not on another coffee break. 2/2

Lukas Trumm reposted

Today TypeScript 5.8 Beta ships the new "erasableSyntaxOnly" flag 🎉 It is designed to pair with Node's built-in TypeScript support, guiding users away from TS-only runtime features such as: ❌ enum ❌ runtime namespace ❌ parameter properties Remember: TS = JS + Types 👍

Rob Palmer@robpalmer.bsky.social

TypeScript 5.8 Beta is released 🎉 🔶 "nodenext" gets require(ESM) 🔶 "nodenext" deprecates Import Assertions 🔷 Checked Conditional Return Types 🔷 erasableSyntaxOnly flag for Node 🔷 Perf: libReplacement flag 🔷 Perf: Program load/update optimizations

I was casually driving home, not really concentrating on what 'someone' was talking about. Suddenly I heard my name! It took me a while to realise that I was listening to none other than Eric Meyer in his Igalia Chats podcast. What an unexpected situation.

I progressively love VueJS 😉 . It was ok in version 2 with Options API, but now on version 3.5 with full Typescript support, Composition API, and latest syntax features it is great.

Having a great time developing a weather related app. I'm using Claude to generate test data and they are hilarious. I've not expected to have fun reading test data 😊

You are up to date.