I use Hermes as the AI agent for my BoFu content work.
My agent has a workspace with the context, instructions, skills, tools, tasks, and previous work it needs to do the job.
I keep every client in the same workspace which lets one agent handle research and content across the agency without setting up a separate agent for every account.
Not everything is automated. While I advocate for more automation, it’s always good to have a human review the work. So after each big step, I look if everything is good or it needs another iteration.
This is what the current setup looks like.
Key takeaways
One workspace holds the instructions, client knowledge, custom skills, and finished BOFU content.
I use Linear for the task queue for both me and the agent. I assign work by adding an
Agentlabel.Three skills cover the main production sequence: find the products, research each product, and write the content.
Hermes can refresh competitor research and use the changes to update existing content, but I review the result before publishing.
Workspace structure
This is a simplified version of the workspace:
workspace/
|-- AGENTS.md
|-- context/
| |-- client-a/
| `-- client-b/
|-- system/
| |-- style-guide.md
| `-- other-instructions.md
|-- .agents/
| `-- skills/
| |-- discover-products/
| |-- research-products/
| `-- write-bofu-content/
`-- work/
|-- client-a/
`-- client-b/AGENTS.md tells Hermes where everything is
My AGENTS.md is mainly a map of the workspace.
It tells Hermes how the folders are organized, where the skills live, where to find the instructions, and which style guide applies to the work. This gives the agent enough information to orient itself before it starts a task.
I do not put every detailed rule in this file. Once it knows where things are, it can read the relevant material for the current job.
context/ holds the knowledge
The context folder contains the information Hermes needs to understand the client and make decisions that fit the account.
That includes reference material, strategy documents, positioning, previous research, competitor information, and anything else that should influence the content.
There is also a shared area for knowledge that applies across clients. Client-specific context stays in a separate folder for each account.
I use one workspace because the harness is smart enough to follow that structure and load the correct client context. The separation is explained in AGENTS.md.
With the rise of Grok Bot and Hermes doing the same, it seems people are going more and more for one agent per role but I haven’t really had the need for this kind of structure yet. I tend to do what works and improve my system only when the need arises.
system/ holds detailed instructions
The system folder contains instructions that apply to the work but do not need to be packaged as skills.
My style guide lives there, along with more detailed rules I haven’t yet converted into skills or are too loose to be skills. AGENTS.md points Hermes to the right instructions instead of trying to contain all of them itself.
.agents/skills/ contains repeatable procedures
Skills are for tasks with a defined method. They tell Hermes how to perform a specific job rather than giving it another collection of general preferences.
I currently have three main skills for BOFU content:
Discover the products in a category
Research each product
Write the content
work/ contains the content
The work folder is where the BoFu content goes.
Each client has a separate folder, just as they do under context/. The context folder contains what Hermes needs to know, the work folder contains what it produces.
That distinction makes it easier for the agent to find previous drafts and update existing content without mixing source material with deliverables.
MCPs and tools I give Hermes
Make.com
Make now lets you create an MCP server and assign workflows to it.
For example, Hermes can request a Google Search Console query through a Make workflow and receive the results. This gives it access to the report without requiring a separate custom integration for every workflow I want to add.
Linear
Linear is where I centralize tasks for myself and the agent.
I write the issues there, decide which ones are ready, and add the Agent label when Hermes can take them. The agent checks Linear, claims eligible issues in Todo, and leaves updates and comments as it works.
That gives me one place to see the work in progress. I do not need a separate agent dashboard (Hermes has a built-in kanban) or a parallel task list that I have to keep synchronized with Linear.
Grok and Firecrawl
Hermes has a web search setting that I can connect to my Grok subscription. In practice, it covers much of the work I would otherwise use Firecrawl for.
Apify
I use the Apify MCP when a site or source cannot be handled through the usual search and scraping route.
Some websites need a specialized scraper. Apify gives the agent another option without making it pretend that it accessed a page it could not retrieve.
If a source remains inaccessible, Hermes reports that it could not access it. It reduces the risk of the AI lying to me.
Why I do not use DataForSEO
I could add a DataForSEO MCP, but I do not find it necessary for this work.
I choose BoFu topics by understanding the category, the competitors, and the questions a buyer asks near a decision. I get a better feel for those opportunities manually than I do from keyword-volume estimates.
For BoFu content, Ahrefs and DataForSEO numbers are guesstimates at best. A narrow query may show little or no volume even when the few people searching for it are valuable.
My three BoFu skills
The workspace tells Hermes where to look, the skills tell it how to complete the work.
1. Finding every product in a category
The discovery skill searches the web for products that belong in a category.
This creates the competitive set for later research. It is useful for obvious competitors, but it also helps surface smaller products that would be easy to miss if I relied only on the names a client already knows.
The goal is to establish which products need investigation before anyone starts writing a listicle or alternatives page.
2. Researching each product
The research skill creates the information needed for an honest comparison.
It checks sources such as product pages, pricing pages, documentation, feature pages, and public reviews. It extracts pricing, features, use cases, limitations, customer opinions, and other facts that may affect the comparison.
The agent records when it cannot access a source. It should not replace missing information with something that sounds plausible.
This research can then support several pieces of content without starting from zero each time.
3. Writing the BoFu content
The writing skill turns the research into the format required by the Linear issue.
That might be a competitor alternatives guide, a best-tools roundup, a versus page, a three-way comparison, a review guide, or one of the other formats I use.
The client strategy documents already explain the positioning and the angles I want to take against different alternatives. Hermes can use those decisions while drafting instead of inventing a new strategy for every article.
I still check whether it applied the right angle. This is one of the areas where the agent can make a reasonable choice that is still wrong for the client.
How a Linear issue becomes BoFu content
The working process currently looks like this:
I create a Linear issue and add the Agent label
|
Hermes finds the issue in Todo and claims it
|
It reads the client context and relevant instructions
|
It discovers or refreshes the products when needed
|
It researches the products and comments on its progress
|
I review the research and correct any gaps
|
Hermes writes the requested BOFU content
|
I check the angle and comparison details
|
I edit and publish the content manuallyThe review after research is important. It is much easier to correct the source material before the draft exists.
The second review is more editorial. I check whether Hermes understood the angle, treated the competitors fairly, and got the small comparison details right.
I publish manually. The agent can prepare the content, but I want the final decision before anything appears on a client site.
Recurring work
The same setup is useful after the first article is finished.
Hermes can check Linear for tasks carrying the Agent label, refresh existing competitor data, and update BoFu content when that research changes.
This matters because competitor content decays. Pricing changes, features move between plans, products change positioning, and reviews reveal new patterns. The original research stays in the workspace, so the agent can compare what it finds now with what it used before.
When an update is needed, Hermes revises the content in work/ and leaves the final review and publication to me.
I’ve become an editor-in-chief
That’s how I feel currently and I like it. I let my agent work and I am responsible for the content quality and strategy.
On top of that, it allows me to spend more time testing content formats to find new ways to rank better on Google and be more visible in AI search.
The agent takes repetitive research and first-draft work off my plate. Each review can also improve the next run. When the angle is wrong, I update the client strategy instructions. When it misses a comparison detail, I adjust the research or writing skill instead of correcting only that article.
This means the workspace gets better as I use it. Hermes starts the next task with the correction already available, while I keep control of the decisions that affect the client.