01What Are Claude Watermarks?

Claude watermarks are invisible markers that Anthropic adds to content generated by new Claude models released after August 2, 2026. This is part of the new EU AI Act. The watermark is hidden inside the text, not a visible stamp.

It is designed to show that AI was involved in creating or editing the content.

If you copy and paste text straight from Claude without changes, the watermark stays. If you use Claude for grammar fixes or edits on a human-written article, the watermark can still be added. Even voice transcriptions made through Claude may include the watermark.

The system is not perfect, and there are many gray areas about exactly when it applies.

02What Keeps a Watermark and What Removes It

Not all content gets a watermark. Short content, like a two-line social media post, is very hard for Claude to mark. Older models before August 2, 2026, like Fable 5, do not have watermarks yet, but Anthropic is trying to apply them retroactively.

  • Long posts copied and pasted untouched keep the watermark.
  • Using Claude for grammar fixes on a draft adds the watermark.
  • Pasting into a document, blog, or email retains the watermark.
  • Heavily rewriting content in your own words removes the watermark.
  • Reformatting or converting a file can also remove the watermark.

Tools to remove watermarks are expected to appear in the future.

03Why You Don’t Need to Panic About Watermarks

The watermark system is not a big problem for most users. Google has already said it does not care if content is AI-generated or not. What matters is quality and whether people like the content.

Social media platforms also focus on quality, not if AI was used.

Gemini has had a similar watermark for a couple of years already. The host, Julian Goldie, does not feel great about watermarks, but he says it is not something to panic about. The main focus should be on making good content.

04Best Models to Use with Hermes and Agents

For setting up Hermes and agents, Julian recommends different models based on your needs:

  • Fast everyday model: DeepSeek V4 Flash – fast, responsive, reasonably priced.
  • Frontier coding model: Quen 3.8 Max – powerful, best for API use if you are willing to pay.
  • Free local model: LFM 2.5, 2.6B – runs on a Mac Studio or normal setup, half decent.
  • Bulk worker: LFM 2.5, 2.6B – good for large tasks.
  • Reviewer or quality controller: Use a different model than your main one, like DeepSeek vs GLM 5.2 or Kimi K3 vs Grok 4.5.

If you have to pick one, go with Quen 3.8 Max. To change the model in Hermes, type Hermes model in your terminal or use the Hermes dashboard.

05How to Clean Up a Messy Obsidian Second Brain

If your Obsidian setup with Graphify and agents has become messy, Julian recommends using Claude’s desktop app to ask it to clean up your Obsidian memory. Use the PARA framework (Projects, Areas, Resources, Archive) to organize notes. Do not ask Claude to delete anything; instead, ask it to put notes into an archive so you keep a backup.

  1. Open Claude’s desktop app.
  2. Provide the documentation of how Obsidian works.
  3. Explain that you want to use the PARA framework.
  4. Ask Claude to organize and filter your notes so agents can pull the right context.
  5. Let Claude gradually fix the links between notes.

This makes your setup cleaner and faster for agents to use.

06Using Agent OS and OmniRoot Remotely

Agent OS is a system where all your agents plug into one place. You can use it remotely with Tailscale on your phone or other devices. OmniRoot runs on a local host, which means it is designed to work on your own computer, not in the cloud.

If you want to connect a local OmniRoot to Tailscale and use it from separate devices, you may need to run OmniRoot on a VPS (virtual private server).

Julian does not use Windows, so he is not an expert on Windows setup. He asked his agent to come up with ideas and provided a guide for Windows users. The key point: OmniRoot is local, so accessing it remotely requires additional steps like using a VPS or WSL experts.

07Tips, Limits, and Mistakes to Avoid

When dealing with watermarks, avoid copying and pasting AI content without changes. Always rewrite heavily if you want to remove the mark. For short posts, you probably do not need to worry.

For Hermes models, do not pick a model that is too slow for your daily use. Kimi K3 is good but currently slow.

When cleaning Obsidian, never ask Claude to delete notes. Archive them instead. This preserves history.

For agent OS setup, do not assume you can run local tools like OmniRoot in the cloud without extra configuration. Each tool has its limits, and you must match the model to the task: speed for everyday, frontier for coding, free for bulk, and a different model for reviewing.

08Who This Information Is For and Why It Matters

This episode is for anyone building AI agents, managing content, or using tools like Claude, Hermes, Obsidian, and Agent OS. If you are a creator, marketer, or developer who wants to stay ahead of AI watermark rules and optimize your workflow, the advice here is practical and current.

Understanding watermarks helps you avoid surprises and keep your content authentic. Knowing which models to use with Hermes saves you time and money. Cleaning up your second brain makes your agents more reliable.

And using Agent OS remotely gives you freedom to manage your team from anywhere. This knowledge matters because it directly affects the quality and efficiency of your AI-powered projects.