Of all the Claude Code agents I’ve built, the one that’s made the most immediate difference to my day-to-day isn’t the one that manages Airtable schemas or gates new projects or fires off n8n workflows. It’s the one that writes client emails.
If you’ve been following the minion army saga on this blog, you know I’ve been building out a roster of Claude Code subagents, each one named after a Minion character and responsible for a specific slice of my workflow. I’ve got agents that write blog posts, build automations, review code, and manage deployments. But the one that changed how my team communicates with clients is A.G.N.E.S.: Approachable, Gracious, and Notably Entertaining Scribe.
She writes client communications. And she is very good at it. (Pssst… I put a link to her generic file towards the bottom, go grab it, its free… I won’t even ask you to register for it!)
What a Claude Code Agent for Client Communications Actually Does
Before I get into Agnes specifically, let me explain what I mean by a Claude Code agent in this context. A Claude Code subagent is a markdown file stored at ~/.claude/agents/ that gives Claude Code a persona, a defined role, a set of instructions, and behavioral guardrails. When you invoke that agent by name in Claude Code, it shows up with that specific personality and expertise loaded in. It doesn’t just use the default assistant behavior. It uses yours.
For client communications, that distinction matters more than you might think.
Most AI writing tools will generate technically correct emails. They’ll hit the key points. They’ll be grammatically clean. But they will also sound like a press release written by a robot who read too many corporate style guides. The tone is sterile. The warmth is performed rather than real. And if you’re in an MSP or a vCIO role like I am, your clients can feel the difference.
Agnes solves that problem because I built her to sound like the best version of a warm, competent, slightly funny IT communicator. Not generic. Not corporate. Specific to the kind of relationship I want my team to have with clients.
Why I Named Her Agnes
In the Minions movies, Agnes is the youngest of the three girls the villain Gru adopts. She’s enthusiastic, genuinely kind, and completely sincere in a way that somehow makes everyone around her feel better about their situation. That felt right for a client communications agent.
A.G.N.E.S. stands for Approachable, Gracious, and Notably Entertaining Scribe. The backronym came after the name, as backronyms usually do, but it actually fits. The goal for every email Agnes writes is to be approachable enough that clients don’t feel talked down to, gracious enough that even difficult conversations land without creating friction, and entertaining enough that people actually read the whole thing instead of skimming to the action item.
That last one is harder than it sounds. Getting someone to read an IT communication all the way through is a genuine skill.
What I Built Into Her Personality
Agnes isn’t just a generic “write client emails” instruction. She has specific behavioral rules baked into her system prompt that shape how she approaches every piece of content.
She knows about Brie and Moose. Brie is my tuxedo cat who has become something of an unofficial IT mascot for one of my clients. Moose is the orange tabby who fills that same role at another. Both of them appear in client-facing content as part of our brand voice, and Agnes knows when and how to work them in without forcing it. That kind of specificity is something a generic AI tool won’t have unless you give it to her explicitly.
She avoids em dashes and hyphen-heavy sentences. (Yes, this is a rule I have for my own writing and yes, I enforced it on my agent. I contain multitudes.) She writes in active voice. She keeps subject lines human and clear rather than clickbaity. She calibrates tone based on the type of email: an update email about a completed project sounds different from a proactive heads-up about a potential issue, and Agnes knows the difference.
She also understands the relationship context. When I’m writing as a vCIO for a nonprofit client, the framing is different than when I’m writing from the MSP side. Agnes holds both versions and I tell her which context I’m working in when I invoke her.
The Build: What Her Agent File Actually Looks Like
An Agnes invocation in Claude Code looks like this:
claude --agent agnes "Write a project completion email for the MFA rollout at [Client]. The rollout covered 474 users. Key outcomes: no major incidents, three helpdesk tickets resolved same day, full adoption in 12 days."She takes that brief and returns a complete email, subject line included, in the right voice and tone. I review it, sometimes adjust a line or two, and send it. The whole process takes about three minutes instead of fifteen.
Her agent file includes:
- A clear role definition: client communications writer for an IT services organization
- Tone rules: warm, direct, humor-forward where appropriate, never condescending
- Formatting rules: no em dashes, active voice, short paragraphs, human subject lines
- Brand context: Brie and Moose, the mascot cats, when to reference them
- Output structure: subject line first, then body, then a note if any section needs human review
- A standing instruction to flag anything that requires specific technical details she doesn’t have
That last one is important. Agnes doesn’t invent facts. If I give her a sparse brief, she’ll write what she can and tell me what she needs to finish the job properly. That’s a behavior I had to explicitly build in because by default, language models will fill gaps with plausible-sounding content. In client communications, plausible-sounding but wrong is a problem.
The Part That Surprised Me
I expected Agnes to save me time. That part worked as planned.
What I didn’t expect was how much she would improve consistency across communications that go out from different people on my team. When a team member needs to send a client update and they’re not sure how to frame it, they can run it through Agnes and get something that sounds like us. Not like them on a rushed Tuesday afternoon. Like us at our best.
That consistency is hard to build manually. It usually takes style guides that nobody reads, feedback cycles that take too long, and a lot of quiet frustration from whoever is receiving the inconsistent communications. Agnes doesn’t fix all of that overnight, but she gives the team a baseline to work from that’s actually good.
What I’m Releasing for Free
Here’s the part where this post becomes useful to you.
I’m releasing a generic version of agnes.md on GitHub so you can pull her into your own Claude Code setup, customize her for your organization’s voice, and start using her without building from scratch. The GitHub version strips out my specific client details and mascot references and replaces them with placeholder variables you fill in for your own context.
You’ll need:
- Claude Code installed and running (the official docs cover setup)
- A
~/.claude/agents/directory - About 20 minutes to customize her voice rules for your organization
The customization is the important part. A generic Agnes is fine. An Agnes who knows your clients, your brand voice, your mascot (optional but recommended), and your communication style is the one who will actually save you time.
Grab the file at GitHub and follow the instructions in the README.
Building Your Own Agent Roster
Agnes is one of 49 agents in my current Claude Code minion army. (Yes, 49. I have a spreadsheet. It’s in Airtable. There is a theme.) She’s one of the easier ones to understand and implement because she has no infrastructure dependencies. She doesn’t connect to a database or fire off API calls. She just writes.
If you’re new to Claude Code agents, Agnes is the right starting point. Get comfortable with how agent files work, how invocation feels in practice, and how to tune a prompt until the output sounds the way you want it to. Then start building the ones that connect to your actual systems.
The agents that connect to real infrastructure, the ones that touch your PSA tickets or your Airtable bases or your n8n workflows, require more careful construction and a lot more testing. Start with the writing agents. They’ll teach you the mechanics without the risk.
I’ll be releasing more generic agents from the roster over the coming weeks, starting with the ones that are most universally useful and least tied to my specific stack. If there’s a particular function you want to see covered, let me know in the comments.
Agnes was a good place to start because she solves a real problem that almost every IT leader I know has: the emails that need to go out, that need to sound human and warm, that keep getting pushed because writing them well takes longer than it should.
She’s got it. Go build yours. I’ll add more soon, I promise
And honestly? My AI minions are now better at client communications than I am before my second cup of coffee. I’m not sure how I feel about that. One step closer to world domination… but first, too many emails.