Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
The only Hermes Agent tutorial you'll ever need...
TL;DR
Alex Finn now recommends Hermes over OpenClaw because it actually stays working — he says OpenClaw has been breaking on “basically every single update,” while Hermes ships fewer but more coherent releases like “Tenacity” that feel curated instead of “the entire kitchen sink.”
Hermes’ big differentiator is self-improvement through skills and memory — every task runs through a skill, and if the skill doesn’t exist or needs refinement, Hermes creates or improves it, so the agent gets better as you use it and exposes the tool-call trail transparently.
His setup advice is simple: use Telegram for messaging and Opus for the model if you can afford it — Finn argues Opus is still the best agent model despite a $300-$400/month API bill, while saying ChatGPT 5.5 is finally usable on a budget and local models like Qwen 36B can work well on hardware like his DGX Spark.
The first two things to do after install are a full ‘brain dump’ and a ‘reverse prompt’ — tell Hermes everything about your work, goals, hobbies, and routines so it builds memory, then ask it what workflows it should run for you based on what it knows.
His beginner and intermediate automations both hinge on cron jobs plus memory — one prompt has Hermes send a daily AI-news briefing with new tools matched to your workflow, and another creates a daily proactive check-in asking for your top priority and suggesting ways to help.
The flashy advanced demo is AI video generation via Hyperframes — when he asks Hermes to make a 30-second short on why people should use Hermes, the agent notices the skill isn’t installed, finds it, installs it, and produces a motion-graphics video in about five minutes.
The Breakdown
Why Hermes suddenly beats OpenClaw
Alex Finn opens with a blunt verdict: Hermes Agent is now better than OpenClaw because OpenClaw keeps breaking. His complaint isn’t subtle — he says he’s spent more time fixing OpenClaw than using it, which is tolerable for him as a technical user but a “major pain in the wazoo” for everyone else.
Reliability, curated releases, and the vibe of the product
What he likes about Hermes is not just uptime, but the feeling that the product is actually thought through. He contrasts OpenClaw’s “kitchen sink” updates with Hermes releases like “Tenacity,” where features such as kanban boards and goals fit a clear theme, and he even calls out the aesthetic as part of why it feels more reliable.
Self-improving skills, transparency, and swarms
Finn highlights Hermes’ self-improvement loop as the killer feature: you ask it to do something, it triggers a skill, and if that skill is missing or weak, Hermes creates or improves it. He loves that you can watch the tool calls and see the skill being refined in real time, and he says the platform is also built for “swarms,” where you can spin up new agent profiles on demand for multi-agent workflows.
Installation is easy; model choice is the real decision
He breezes through setup — paste a command into the terminal, choose your messaging layer and model, done. His strongest recommendation is Telegram because the formatting, approvals, and agent interaction are smooth there, and for models he still backs Opus as the best for reliability, taste, and UI-building, while saying ChatGPT 5.5 is now decent enough for people on the $20 plan after 5.4 was “completely useless.”
Cloud vs local: his practical split
Finn also makes room for local models, especially if you’ve been buying GPUs like he’s been urging people to do. He shows Hermes running on Qwen 36B on his DGX Spark and says the smart move is to run two agents — one local for cheaper, simpler tasks and one cloud-powered for harder, time-sensitive work.
The first move: brain dump everything about yourself
Before any fancy automation, he says every user should start with a brain dump. He demonstrates by telling Hermes who he is — Alex Finn, entrepreneur, 200,000 YouTube subscribers, 450,000 on X, founder of Henry Intelligent Machines, AI-obsessed, Boston sports fan, live streams Monday/Wednesday/Friday at 11 a.m. Pacific — so the system can build durable memory about him.
Reverse prompting: let the agent tell you what to automate
His second must-do is “reverse prompting,” which he pitches with real evangelist energy. Instead of asking the agent to do a random task, ask it what workflows it should create based on what it knows about you; his point is that too many people copy generic use cases instead of using the agent’s memory to surface custom, high-value automations.
Three use cases, from daily tool discovery to AI video generation
The beginner use case is a daily cron job that sends AI news plus a new tool recommendation tailored to your workflows. The intermediate use case is a proactive morning check-in where Hermes asks for your top priority, updates its memory, and suggests ways to help; then he closes with the flashy one: using Hyperframes to generate a 30-second Hermes promo video, where the agent notices the skill is missing, installs it itself, and produces a polished motion-graphics draft in roughly five minutes — not final-post ready, in his words, but good enough to save “thousands of dollars” on editing and prove the gatekeeping is gone.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
