Guido Schmitz

wrote something nice teapawt.net/guidsen/status/2013596…

turn your hermes agent into grok bot theme with @NousResearch github.com/guidsen/hermes-gr…

Mannnn OpenClaw is causing to many headaches. Migrating all our business agents to Hermes @NousResearch now as well. It seems after the acquisition things really went downhill lol

We’ve built 14 companies using AI in 5 days… What happens when you lock up 30 founders & devs in the Morocco dessert for 5 days? 25 concepts built. 14 companies started. Billions of tokens used. Such a great experience! You can just build things 🤝

Man optimizing queries was always sooo time consuming. Really love these frontier models

People sometimes ask how I use AI to automate so much operational work. The honest answer: I don’t have one magical prompt. I started with prompting, like everyone else. But over time, the unit of work kept getting bigger: Level 1: Prompt engineering What do I ask the model? Level 2: Context engineering What information does it need to do the work properly? Level 3: Workflow engineering Which tools can it use? What actions can it take? How do we verify the result? Level 4: Loop engineering What triggers the workflow? What state carries over? When should it retry, stop or ask for help? Loop engineering sounds more innovative than it is. In engineering, we’ve had jobs running at set intervals forever. We call them cronjobs. Nothing new there, haha. But a cronjob is only the trigger. The interesting part is what wakes up when the timer fires. If you want to experiment with this, don’t start by building an “autonomous AI employee.” Start with one boring process that repeats. A few ideas: 1/ Every morning, summarize deployments, incidents, failed builds and cloud-cost anomalies, then flag what needs attention. 2/ Every morning, combine signups, activation, funnel conversion and paid acquisition data into one growth report, then investigate meaningful changes. 3/ Every hour, triage new Datadog errors, connect them to recent releases and prepare tickets or draft PRs. 4/ When CI fails, inspect the logs, reproduce the issue, attempt a fix in an isolated branch, rerun the tests and open a draft PR if everything passes. 5/ When a target account shows a useful signal, enrich it from approved sources, score the fit, update the CRM and prepare personalized outreach for review. 6/ Give a landing-page agent a campaign brief, let it build through the Webflow MCP, run performance and accessibility checks, and return a preview for approval. Start in report-only mode. Once the output becomes reliable, let the loop prepare drafts. Then allow low-risk, reversible actions. Keep approval gates around anything sensitive. A useful loop needs more than a schedule. It needs context, tools, persistent state, a way to reject bad output and a clear stopping or escalation condition. Otherwise, you haven’t automated a workflow. You’ve scheduled a prompt. The goal isn’t to remove human judgment. It’s to remove the need for a human to remember every step, move information between systems and keep pressing “go.”

Hermes Agent by @NousResearch, but make it Grok Bot Get the theme at github.com/guidsen/hermes-gr…

@HermesWatcher 👀

Can Grok Bot fans tell me why that over Hermes where you own your workflows and have model flexibility. I really believe in this AI Sovereignty thing.. Just install Hermes Agent on a $10/m Hetzner VPS where you own all the files, choose different models etc. Then for coding run `claude rc` via tmux and you’ll never need something else lol.

wish @OpenRouter was easily available to host models in Europe. really important for enterprises to adopt AI models in Europe.. any alternatives/tips?

If you're spinning up to many terminals and sometimes forget what's in progress / waiting for you, this /recap skill is for you

having a status on claude code convos would be great. drowning in them lol. when working in parallel (4+ terminals) its hard to understand which ones are fully done and which ones you’re going back in forth with @claudeai cc @bcherny