AaronCQL

Everybody and their mums building the infra. So who's building the damn apps?

miniclaw > claude code channels confirmed

Once telegram markdown and slash commands (/compact, /loop) are supported, it might be over for all claw products.

> the skills repo you think you need: skillsmp.com > the skills repo you actually need: roadmap.sh

@xianxlb Yea it's lengthy and filled with custom instructions I had to skip to read the actual plan. And because the instructions are so intertwined, if something wrong is planned, multiple places need to be changed. 🤔

@AaronCQL i tried superpowers too and my verdict is similar to yours it’s super lengthy and still makes mistakes

Tried out the Superpowers plugin on Claude as it has 97K stars on GitHub. Planned a real-world feature on an existing codebase and compared it to vanilla Claude. Very surprised to see vanilla wins almost hands down: closer to what I wanted and easier to tweak. Am I missing smt??

Back to Gemini we go

@ASvanevik Wanna start a better one together?

Everything in Singapore is modern and efficient. Except the banks. Wow.

Not having implicit script fallback is prolly the better move long term tbh. Short term pain from people shifting over from existing codebases, but feels like the cognitive overload is definitely much less in the long term. I for one welcome this change, it's one of my main gripes using pnpm.

One thing that I know some people will complain about, and @theo did in fact complained about, is how ‘vp dev’ is a built-in command rather than running the dev script like a package manager does. Fundamentally this is because vp is not a package manager or just a task runner - it’s a full toolchain. A package manager doesn’t have a built-in dev feature, but vp does. Running scripts is something vp can do, but you make the intention explicit via ‘vp run’ instead of expecting implicit fallback. I actually very much dislike the implicit script fallback behavior because it forces you to remember what built-in commands the runner has - for example, you can’t run a script named ‘deploy’ with ’pnpm deploy’ because… turns out pnpm has a built-in deploy command. With the large command surface vp has, implicit fallback is only going to be confusing than helpful.

@jestermolecule Appreciate it! Starring the repo is more than enough for now ☺️

@AaronCQL would you be open to receiving donations to ur github for this?

Just formally released v0.1.0 of miniclaw: a minimal Telegram agent powered by Claude Code, designed to be self-modifiable. Telegram still remains my favourite chat UX on both mobile and desktop. Threads, file sharing, voice messages, real-time notifications - why reinvent the wheel when the perfect interface already exists? Using Claude Code as the harness means all improvements by Anthropic get captured immediately with zero maintenance work, and you get to enjoy a familiar harness that you probably already use. It's built on top of Claude's auto-memory system: no complex vector dbs, embeddings, or bloated background services. Just plain markdown files you or Claude can easily read and edit. Lastly, the codebase is *500x smaller* than OpenClaw - small enough for you and Claude to understand in one go. Want a new feature? Just ask miniclaw to modify itself for you.

If you're a fairly technical person and have never built your own server rig or self-hosted stuff you're currently paying for, here's your sign to start. All you need is a $100 CPU and you can get your own dedicated ad blocker, VPN, media server, file sharing, and AI agent.

Exceptionally relevant to myself

@JupDevRel The real alpha is always in the comments

@AaronCQL miniclaw and jup cli goes wild

@sssionggg Why is man frontrunning company secrets

if you need a Jupiter CLI for your claw bots, check this out: github.com/jup-ag/cli

That's why I'm still on Claude Code, betting on it to outlast all other harnesses in the long run since it's literally Anthropic's bread and butter. Sticking to one that will constantly get updates and being able to deeply understand/tweak it will be more powerful than any synthetic benchmarks imo. Similar philosophy to programming langs.

@AaronCQL Cos I have not had time to try it out yet. Probably would try at some point. Too much shit to try, dont even know which one is worth trying lol

The agentic harness space is getting wild. New players keep popping up: Deep Agents (LangChain), Hermes Agent (Nous Research), OpenCode (120k+ stars??), Goose, ForgeCode... Seems like the LLM matters less now. It's the harness that differentiates: context management, tool dispatch, memory, safety. Which harness are you using and why?

@xianxlb have you tried jupiter global?

dear onramp providers, why do you not support SGD? sincerely, a solo vibecoder

@tero0x I'm on vanilla Claude Code. Question is why are you not using Forge? LOL

@AaronCQL Using opencode mostly with superpowers plugin. Can check out tbench.ai/leaderboard/te…. Got comparisons between these harnesses - some of them game the benchmark though. What do you use?

@beeman_nl Been telling folks cloudflare is a way better product for years

Friends don't let friends use Vercel 🤷‍♂️