Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
Hermes Agent is blowing me away...
TL;DR
Alex Finn now recommends Hermes over OpenClaw — his core reason is boring but decisive: Hermes updates are more reliable, while OpenClaw has repeatedly broken after updates and cost him “20 minutes fixing it” instead of actually using it.
Hermes’ edge is self-improvement plus memory — Finn shows how every task runs through a skill, then improves or creates that skill if needed, so the more you use Hermes on things like YouTube transcripts, the better it gets over time.
The best setup, in his view, is Telegram + Opus — he calls Telegram “king for AI agents” because of approvals and tool-call visibility, and says Anthropic Opus is still the strongest agent model even if it can run $300-$400/month.
If you’re budget-conscious, Finn still gives you options — he says ChatGPT 5.5 is finally usable for agents after 5.4 was “completely useless,” and local models like Qwen 36 on his DGX Spark can handle coding and simple tasks for near-zero ongoing cost.
His two must-do onboarding steps are ‘brain dump’ and ‘reverse prompt’ — first tell Hermes everything about your career, goals, hobbies, and workflows, then ask it what automations and workflows it should build for you based on what it knows.
The practical payoff is proactive automation, not just chat — his three examples are a daily AI-tool recommendation cron job, a morning priority check-in that updates memory and suggests help, and a Hyperframes-generated 30-second promo video made in about five minutes.
The Breakdown
Why Hermes suddenly beats OpenClaw
Finn opens with a blunt take: Hermes Agent is now better than OpenClaw because OpenClaw keeps breaking. As a technical user, he can patch things fast, but he says normal users shouldn’t have to spend time fixing an agent after every update just to get work done.
Reliability, curated releases, and the self-improving loop
What wins him over is that Hermes ships less often but with more discipline — releases have themes like “Tenacity,” and the features actually fit together instead of feeling like “the entire kitchen sink.” He also highlights Hermes’ self-improvement loop: every request triggers a skill, and if the skill is missing or weak, Hermes updates or creates it, so repeated use literally makes the system better.
Built for tinkerers, local models, and swarms
Finn clearly loves that the New Research team feels like “hackers,” with support for local models, LoRAs, and machine learning experimentation. He also calls out swarms as a major differentiator: you can just say “build a new Hermes profile” and spin up another agent fast, which fits his multi-agent workflows.
Installation is easy; the real choices are Telegram and model selection
He breezes through setup as basically copy-paste into terminal, then focuses on the two choices that matter: how you talk to the agent and which model powers it. His recommendation is strong and simple — use Telegram because approvals and formatting are great there, and use Opus if you can afford it because it has the best “taste,” UI output, and reliability for agent work.
Opus if you can pay, ChatGPT 5.5 or local if you can’t
Finn pushes back on sticker shock around spending $300-$400 a month on Opus, framing it as an investment in yourself rather than a distraction expense like Netflix or Game Pass. Still, he gives alternatives: ChatGPT 5.5 is now “actually usable” after 5.4 flopped badly for agent tasks, and local setups like his Qwen 36 running on a DGX Spark can do strong coding work without ongoing API spend.
The first thing to do: tell Hermes who you are
Before any fancy workflows, he says to “brain dump” your entire life into the agent — career, goals, audience, hobbies, even stuff like his Boston sports fandom and 200,000 YouTube subscribers. The point is to feed Hermes’ memory system enough detail that every later recommendation or automation is personalized.
Reverse prompting: let the agent tell you what to automate
This is one of Finn’s favorite ideas: instead of arriving with polished use cases, ask Hermes what it thinks you should be doing based on what it knows about you. His pitch is that people get stuck because they copy generic use cases, when the better move is to have the agent surface custom workflows tailored to your actual work and ambitions.
Three use cases: daily tools, proactive check-ins, and AI video generation
His beginner example is a daily cron job that scans AI news and recommends new tools based on your workflows and prior usage. The intermediate one is a morning check-in that asks for your top priority, updates memory, and suggests how Hermes can help; the advanced one uses Hyperframes to generate a 30-second motion-graphics short, which Finn says took about five minutes and could save “thousands of dollars” versus hiring an editor or team.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
