@aarondfrancis you've created something beautiful with the latest Solo update. The MCP, timers, todos and scratchpad combo unlocks so much potential. Pretty crazy to see agents communicate with each other.
Been playing around with Superpowers. Highly recommend the brainstorming skill. It produces really good clarifying questions. Results in good plans and sub agent implementation is chefs kiss. Getting best results with Claude.
https://t.co/drUuxYoeHs
If i pay for usage it should be up to me how I spend that usage. Whether thats on Claw or something else. Basically how it is with any normal currency. Their tokens are basically food stamps now and you can only spend it with them.
Using their Dispatch tool or whatever the rival Claw features are called is basically the same as Claw. The token use for the work being done is not different.
However, if this helps Anthropic to make their products better, I’m all for it. Im (still) an happy customer. In the end it’s also their product, their rules.
Spent 4 months jumping on every new AI tool that dropped. Now my system finally starts to click.
Using Claude for planning, ChatGPT for coding (it's really fast and good). Gemini and Amp are used for feedback with @aarondfrancis counselors.
Using @usemonologue for speech to text. Don't use its local model for transcription, it's really slow and makes Monologue feel like a subpar product. It really isn't.
Any coding task I work on goes through @openclaw (works best with ChatGPT):
- Discuss what needs to get done on Telegram/Slack
- Complex task? It creates a @claudeai session in @getpolyscope, provides me a link to it, and I work on the plan there. Polyscope makes it easy to inspect and comment on the plan
- Kick off @aarondfrancis counselors for a review
- Plan done, OpenClaw takes over and drives Polyscope for me. Meanwhile I do other work
- OpenClaw pings when work is done
- I review the work (code + UI) in Polyscope and request adjustments where needed
- All good? OpenClaw pushes the changes and handles any pre-commit/push issues
All work is being done outside my laptop now. Started automating midnight improvements and maintenance through OpenClaw cronjobs, same flow as above. In the morning I review the Polyscope sessions it created.
Biggest realization was how well OpenClaw can discover API endpoints and CLIs and function as a central router to connect tools and automate workflows. When I let it use a new tool like Polyscope, I always let it write a skill first. Memory still sucks for me with OpenClaw but having dedicated files for it to read and write to for specific tool use works really well. OpenClaw really helps with automating the boring, repetitive parts of my work.
I only use one OpenClaw agent now. One entrypoint is really convenient. All work is done through sub-agents so the main thread stays responsive. When I need to work on a specific project I switch to Slack. Each project has its own channel which helps keep things organized.
Used @opencode mainly before Polyscope (still big OpenCode fan), but not having access to Claude made me reach for it outside OpenCode. It really helps to stick to one tool per primitive/category (STT, Agent IDE, Hosting, etc). Made things a lot simpler and more manageable. Haven't touched Claude Code, Codex, Soloterm and Cursor in a week now.
I also had to adjust to the idea of having a fluid toolchain. A tool I use today can be broken or nerfed tomorrow, or a more capable one gets released. I use one tool per category but I'm no longer attached to the tool itself. Tools within each category work pretty much the same and OpenClaw handles configuration for the most part, so swapping is quite effortless.
I sure fell into the trap of jumping on every new tool that came out and spending way too much time and money on them. Yet, it got me to work on a system. This is an open door, but having a system is more important than ever, without one you'll get lost in the AI jungle we're currently in.
Next week we’ll be at @LaraconEU with @freekmurze, @mattiasgeniar and @nckrtl.
And we have one full conference ticket to give away! 🎁
So it’s raffle time: if you’re an Oh Dear client and can attend the conference next week Monday and Tuesday, reply here, we’ll pick a random winner!
@burkeholland@github That is really cool indeed, but isn’t this how agents are designed? For each agent definition you can specify the model explicitly. Or do you mean it spawned the agents ad-hoc with the requested models without being specified up front in your config?
Anyone I know has experience with AI image generation based on lora/flux with something like https://t.co/KckurqCcGJ? I have quite some reference material 40+ images (hand drawn illustrations) in high quality but the trained set doesn't produce anything that comes near the reference material. Midjourney moodboards same story :/
Oh my days. @openclaw just figured out how to use @opencode api to create new sessions. Sessions appear live in the opencode UI. This opens up so many cool automation/orchestration workflows 🤯
@opencode's success seems making it more unstable by the day. Switched to @conductor_build for local development. What a beautiful app. Wish they had had external server support like opencode.
Also shows its worth to keep llm configs in sync with each other so you can easily switch.