How to Run AI Systems with Claude Code and Hermes

TL;DR

Use Claude Code to build a system once, then hand it to Hermes for scheduled, ongoing operation. Connect both tools to the same memory account and ID so they share context, retain corrections, and improve workflows without repeated explanations. Reliability improves when systems are designed for handoff, verify their own results, and use cheaper models for routine operation.

Transcript

Last night, I was asleep, but my AI agent was not. It went through my whole inbox, and it closed a $6,000 client by the time I woke up. I didn't have to lift a finger. Here's how. I run two AI tools together, and on their own, each is just a tool. Clot Code is the best builder I've ever used. Hermes is the best operator. But put them together and t... Read More

Key Insights

• Claude Code is positioned as a builder that creates apps, scripts, agents, websites, services, and infrastructure during interactive work sessions. Its main role in the proposed workflow is constructing or upgrading the machine that another tool will operate repeatedly.
• Hermes is positioned as an operator that can live on a server or personal machine, run scheduled jobs, receive instructions from a phone, and continue working in the background. Its strength is operating existing systems rather than serving as the primary coding environment.
• The central workflow is to build once with Claude Code and let Hermes run the resulting system continuously. Although both tools have overlapping capabilities, assigning each one the job it is designed for creates a clearer division between development and ongoing operations.
• Shared memory works by pointing both tools to the same memory account under the same identifier. This additive layer lets either tool write information that the other can later retrieve without replacing the separate memory capabilities already present in Claude Code or Hermes.
• Mem0 is the presenter's chosen shared memory option because it is described as fast, free to start, and easy to configure. The transcript also names Zep for time-aware facts, Letta for auditable memory, and Supermemory for simple, fully local storage.
• Performance feedback can improve reusable workflows over time. In the content example, Hermes examines second-by-second retention, audience drop-offs, editing cuts, delivery energy, and rambling, then produces a brief that can be fed back into Claude Code to refine the editing skill.
• Corrections can become persistent operational instructions when Hermes analyzes a failure, identifies its cause, and rewrites its own guidance. The intended result is a system that retains lessons from client emails, broken pages, declining metrics, and other problems instead of repeating them.
• Cost and reliability depend on deliberate implementation choices. The presenter recommends using an expensive model for building, a cheap or local model for daily operation, designing the system for Hermes from the beginning, and requiring every job to verify its result before completion.

Questions & Answers

Q: How do Claude Code and Hermes work together?

Claude Code builds the underlying machine, such as a website, backend, script, agent, editing skill, service, or business system. Hermes then operates that machine on a schedule or in the background, monitors its performance, catches problems, and reports needed changes. When a substantial upgrade is required, the workflow returns to Claude Code before handing the improved system back to Hermes.

Q: What is the difference between an AI builder and an AI operator?

An AI builder works interactively with the user to create or upgrade software and workflows, usually in focused bursts at a keyboard. An AI operator runs those completed systems repeatedly, often on a server or personal machine, and can work on a schedule in the background. The transcript assigns Claude Code the builder role and Hermes the operator role.

Q: How can Claude Code and Hermes share the same memory?

Connect both tools to one shared memory service, using the same account and identifier for each connection. The memory layer acts like a notebook that both tools can read and write. Information saved through Claude Code can then be retrieved through Hermes, and the reverse also works. This shared layer adds to their existing memory systems rather than replacing them.

Q: Which shared memory tools can connect Claude Code and Hermes?

The transcript identifies Mem0, Zep, Letta, and Supermemory as possible shared memory options. Mem0 is the presenter's personal choice because it is described as fast, free to start, and easy to run. Zep tracks how facts change over time, Letta exposes editable memory for auditing, and Supermemory offers a simple local option where data remains on the user's machine.

Q: How can an AI operator improve a content workflow?

Hermes can monitor more than total views by examining second-by-second audience retention, the moments when viewers leave, cuts that lose attention, changes in speaking energy, and sections containing excessive rambling. It can turn those observations into a brief for the next production. Feeding that brief back to Claude Code allows the original video editing skill to improve after each published video.

Q: How can this AI setup support client operations?

Claude Code can build the websites, systems, services, or products delivered to clients, while Hermes monitors the operational layer after launch. Hermes can keep track of each client, identify details that might otherwise be missed, and flag developing problems early. The intended division is that Claude Code creates the service infrastructure and Hermes helps keep it running while supporting client satisfaction.

Q: How does the system improve after failures or corrections?

When something goes wrong, such as a missed client email, a broken page, or a declining metric, Hermes can examine what happened, determine why it occurred, and rewrite its instructions. A user correction can therefore become a persistent skill or operating rule. Shared memory preserves those lessons, while Claude Code remains available when the system needs a larger upgrade or a new feature.

Q: How can Claude Code and Hermes reduce operating costs and failures?

Use the strongest and more expensive model when Claude Code is building the system, then let Hermes handle recurring work with a cheaper or local model. Design the system for Hermes to operate from the beginning instead of improvising the handoff later. Add an explicit instruction requiring each job to verify its result before declaring completion, which the presenter says prevents many silent failures.

Summary & Key Takeaways

• Claude Code serves as the builder for websites, scripts, agents, editing workflows, and business infrastructure. Hermes serves as the operator that runs those systems continuously, monitors results, identifies problems, and reports needed changes. The core workflow is to build a machine once, then hand it to an operator for repeated execution.

• The setup supports website monitoring, content improvement, and client operations. Hermes can watch signups, email open rates, audience retention, communication patterns, and client activity. Its findings can guide changes to existing systems, including improvements to a video editing skill originally built with Claude Code and refined through performance feedback.

• A shared memory layer lets Claude Code and Hermes read and write context through the same account and identifier. Tools mentioned include Mem0, Zep, Letta, and Supermemory. The recommended operating practices include designing for handoff, verifying results, using cheaper models for routine jobs, and preserving corrections as reusable instructions.