Hermes Agent for Content Production: UK Business Guide (2026)

The pattern Ampliflow uses for our own content publishing — and the one we deploy for UK SME content marketing teams — is Hermes drafts, Claude polishes. Hermes runs on a content calendar, drafts long-form articles + social variants + visual briefs at 03:00 London time, drops them in a review queue. The marketing lead reviews, polishes, publishes. The agent handles the cold-start problem; the human handles voice, judgement, and final quality. This guide covers the workflow, the skill files that make it repeatable, the cost economics vs alternatives, and the discipline that prevents shipping AI slop.

Last updated: May 2026 · Covers Hermes Agent v0.13 + the content production pattern we run for ourselves and Ampliflow clients

• Hermes drafts at 03:00 London on a content calendar; humans polish during the working day
• Total monthly cost for a 4-8 article pipeline: £40-80 in model spend + £0-4 server cost
• The discipline that makes this work: Hermes drafts always go to a human reviewer, never directly to publish
• The pattern fits agencies, in-house marketing teams, and founder-led businesses with content needs
• Pairs naturally with Hermes Agent for sales/CRM automation and Claude Code for content production

The pattern in action

A scheduled Hermes job runs at 03:00 London time, Monday through Friday. For each item on the content calendar:

1. Reads the brief from /home/ubuntu/content-briefs/YYYY-MM-DD-slug.md
2. Pulls relevant context from internal sources (GSC top queries, GA4 top pages, prior published content, customer testimonials)
3. Drafts a 2,000-3,000 word article in the documented house voice
4. Generates supporting visuals briefs (Mermaid diagrams, chart data, hero image prompts)
5. Saves the draft to /home/ubuntu/content-drafts/YYYY-MM-DD-slug.md
6. Sends a WhatsApp message to the marketing lead: "Draft ready for {slug}. ~{word count} words. Brief queue: {N} items remaining."

By 09:00 the marketing lead reviews the draft on phone over coffee, accepts the structure, and books polish time later in the day. By end-of-week, four to eight articles are polished, hero images generated, internal links wired, and queued for publish.

The cold-start problem — staring at a blank document at 16:00 on a Tuesday — is gone. The hard work (voice, judgement, fact-checking, internal linking) is preserved for the human. The grunt work (research, structure, first-pass prose) is automated.

Why Hermes specifically

Three properties make Hermes the right tool for content production at SME scale:

Self-hosted, no per-piece pricing. Commercial alternatives (Jasper, Copy.ai, Writer.com) charge per word, per seat, or per project. A Hermes deployment costs the underlying API tokens (£40-80/month for 4-8 long-form pieces) plus your server (£0 on Oracle Free Tier, ~£4/month on Hetzner). Total monthly cost is lower than a single seat on most commercial platforms.

Persistent context. Hermes' memory layer accumulates your brand voice, your client tone, your topic expertise over time. By month three, the agent's drafts sound like your team writes — because the agent has seen your team's polish patterns and learned from them. Commercial platforms reset every project.

Multi-channel output, no extra integrations. Hermes can draft an article, generate matching social posts, write a newsletter version, and produce briefs for video content — all in one workflow, all routed to the same review queue. The pattern is documented in our Hermes use cases overview.

For agencies running content for multiple clients, the same Hermes deployment can serve multiple brands with separate skill files per brand voice — much cheaper than per-client commercial seats.

The skill file that makes it repeatable

Drop this in `~/.hermes/skills/draft-article/SKILL.md` (adapted to your stack):

`yaml

name: draft-article description: Drafts a long-form article from a brief in the documented house voice. Use when the user asks to draft an article, write a blog post, or process a content brief.

Inputs

• Brief file at /home/ubuntu/content-briefs/{slug}.md containing:
• Topic
• Target keyword (verified from DataForSEO research)
• Target audience
• Internal links to weave in
• Word count target

Process

1. Read the brief file
2. Pull recent context:

• Top GSC queries for this topic (via GSC API)
• Last 3 published articles in adjacent topic clusters
• Any customer testimonials matching the topic

1. Draft the article following voice rules in /home/ubuntu/voice/HOUSE-VOICE.md
2. Apply structural rules:

• 60-80 word AEO answer block at top
• 4-5 bullet TL;DR
• One-line answer at start of every H2
• Frequently Asked Questions section with 6-8 Q&As
• Related Reading section with 3-5 internal links

1. Generate visual briefs at /home/ubuntu/content-drafts/visuals/{slug}.md:

• Hero image prompt (Dan Koe pen-and-ink style)
• Mermaid diagram suggestions
• Chart data if relevant

1. Save draft to /home/ubuntu/content-drafts/{slug}.md
2. Send WhatsApp confirmation: "Draft ready for {slug}, ~{word_count} words"

Forbidden

• Don't publish anywhere directly (drafts only)
• Don't fabricate statistics or quotes (mark with [VERIFY] if uncertain)
• Don't use AI tells: em-dash overuse, "genuinely", "absolutely", breathless openers
• Don't claim Ampliflow has done things we haven't done

`

The 30-line skill plus a documented HOUSE-VOICE.md is the entire setup. After the first month of refinement (every typo and tone miss becomes a rule in HOUSE-VOICE.md), the drafts become reliable.

What it costs vs alternatives

At our typical cadence (4-8 articles/month at 2,000-3,000 words each):

The economics favour Hermes for any team producing more than 2-3 long-form pieces monthly. Below that volume, ChatGPT Plus + manual is fine. Above 8 pieces monthly, the gap widens further.

Where commercial tools win: when you need a polished UI for non-technical contributors, or when you need pre-built integrations to specific platforms (HubSpot, Mailchimp templates, etc.). Hermes is the power layer; commercial tools are the friendly UI.

The discipline that prevents AI slop

Three rules we follow every time:

1. Drafts always go to a human reviewer

No exceptions. The agent's WhatsApp confirmation is "draft ready," not "published." Every piece passes through human review for voice, fact-checking, and judgement before going to the CMS publish step. This is non-negotiable — the moment you start direct-publishing AI drafts, you ship slop.

2. AI-tells get caught at polish

Hermes drafts have a recognisable pattern: em-dash overuse, "genuinely" / "absolutely" filler, breathless openers ("In today's rapidly evolving digital landscape..."), over-qualified parentheticals. The polish pass strips these. Maintain a documented list (we have one at ~/voice/AI-TELLS-TO-AVOID.md) and grep every draft against it.

3. Every fact gets verified before publish

The agent will cite real-sounding statistics that don't exist. The agent will paraphrase a quote from a source that doesn't say what it's paraphrased to say. The polish pass treats every numerical claim, every named company, every quoted source as suspect until verified against the original. This is the slowest part of the polish pass and the most important.

If your content is going to be cited by AI search engines (ChatGPT, Perplexity, Gemini), this matters more — those engines will pick up your statistics and amplify them. Bad data spreads.

How Ampliflow runs this for our own content

The 23-article authority push that produced this article used the pattern. Hermes drafted overnight, Claude (via this terminal session) polished and shipped during the day. Total Hermes-side time per article: roughly 90 seconds of model invocation. Total polish time: 30-90 minutes per article depending on complexity.

The economics for the push: ~£40 in Hermes model costs + Claude Code subscription cost (already running for engineering work) + 30-40 hours of polish time. A freelance writer producing 23 articles of comparable depth would have charged £4,600 (£200/article) to £11,500 (£500/article).

The pattern compounds: the next 23 articles in the cluster will draft faster because Hermes has learned more of the voice + the topic clusters. By the time the cluster is complete (50-100 articles), the marginal cost per article drops further.

When NOT to use this pattern

Three contexts where direct LLM use beats the Hermes pattern:

Single-article one-off

If you're writing one article — not a content programme — the operational overhead of setting up Hermes isn't worth it. Use Claude or ChatGPT directly, polish, ship.

Highly bespoke / strategic content

Founder thought-leadership pieces, anchor case studies, anything that depends on first-person voice — write it yourself or with light AI assistance. The Hermes draft-and-polish pattern works for evergreen, research-heavy, format-consistent content. Strategic pieces aren't that.

Compliance-heavy content (regulated sectors)

For FCA-regulated firms, NHS suppliers, legal advice content — the human-in-loop needs to be deeper than "polish pass." Each draft should be reviewed by a subject-matter expert against compliance requirements. Hermes can still help with structure + research, but the human review is the long pole, not Hermes' speed.

Frequently asked questions

Does Hermes-drafted content rank in Google?

Yes, with proper polish + AEO formatting. Google's published helpful content guidelines explicitly state that "AI assistance" is fine when the output is genuinely useful + has human accountability. The 23-article authority push that includes this article is itself a real-world test — early data shows the articles indexing within 2-7 days of publish, comparable to manually-written content from the same domain.

Can Hermes write in different brand voices for different clients?

Yes. Use separate skill files per brand: draft-article-brand-a/, draft-article-brand-b/. Each skill points to its own HOUSE-VOICE.md file. The agent reads the right voice file based on which skill is invoked. This makes Hermes particularly useful for agencies running content for multiple clients.

What's the polish-to-draft time ratio?

Roughly 30-90 minutes of polish per 2,000-3,000 word draft, depending on complexity. Comparison-heavy articles take longer (more verification); evergreen tutorials take less. Practically, one polish hour produces one shippable article.

How do I prevent Hermes from making things up?

The skill prompt should include "Don't fabricate statistics or quotes; mark with [VERIFY] if uncertain." During polish, grep for [VERIFY] and either confirm against a source or remove the claim. Over time, your skill prompts get more specific about what's verifiable from your internal sources vs what needs external checking.

Can Hermes draft in different formats (long-form, short-form, email, social)?

Yes. One skill per format — draft-article, draft-social-thread, draft-newsletter, draft-video-script. Each skill has its own structural rules. The agent picks the right skill based on the brief.

What happens to my content if Hermes drafts something problematic (libellous, biased, factually wrong)?

The polish step is the gate. The agent's draft is liability-free until you publish it. Treat every draft as you'd treat copy from a junior writer — review for accuracy, tone, and editorial quality before it ships. The legal accountability is on the publisher, not the drafter.

Does Hermes handle SEO research?

Hermes can pull data from APIs (DataForSEO, Ahrefs, GSC) if you wire those integrations. But the strategic SEO decisions (which keywords to target, what topic clusters to build) should stay human — those are calls that compound over years.

How does this compare to using Claude Code for content?

Claude Code is the engineering-side equivalent — it lives on a developer's laptop, writes content interactively, edits in place. Hermes is the operational equivalent — runs unattended on a server, drafts on a schedule, routes outputs to channels. For a content marketing team, Hermes is usually the better fit; for a single content founder who's already comfortable in the terminal, Claude Code works too. The deeper comparison is in our Claude Code for non-developers guide.

Related reading

• ↑ What is Hermes Agent? A UK Business Guide — the foundational pillar
• ↔ Hermes Agent — Real Business Use Cases — the use cases pillar; this content production pattern is Use Case 3 in detail
• ↔ How to Deploy Hermes Agent — UK Business Complete Guide — required for any of these workflows
• ↔ Hermes Agent Skills — Build Custom Extensions — the skill-writing patterns this workflow depends on
• ↔ AI Content Marketing Tools UK 2026 — the broader UK content marketing tooling stack Hermes fits inside
• ↔ Claude Code for Non-Developers — the laptop-based equivalent for solo content founders

What should you do next?

The pattern takes about a day to set up the first skill + voice file, and another month of refinement to dial in the voice. After that, it runs itself.

See Hermes-powered content automations we run for clients →

Or to scope your specific content workflow: Book a free Hermes use cases review →