AIMM Session — August 28, 2025: Claude Skills and the Maintenance Cost You're Not Calculating

“Prompting is the new literacy for knowledge entrepreneurs — not a technical skill, but a strategic capability.” — Lou’s AIM Writing Team skill, generated live during the session

30-Second Summary

This session centered on Claude Skills, Anthropic’s newly released format for packaging and deploying reusable AI capabilities as portable zip files. Lou walked the group through building a complete multi-role writing team — researcher, strategist, writer, editor, publisher, orchestrator — all housed in a single skill with a self-evolving memory layer. The conversation then widened into a sharp discussion about when to adopt new tools at all, landing on a principle that applies well beyond AI: factor in the full maintenance cost before you commit.

Claude Skills: Build Once, Deploy Everywhere (Sort Of)

The announcement that went viral last week is, at its core, a prompt in a zip file — and that’s not a knock against it. Lou opened by laying out what Claude Skills actually are: a folder structure anchored by a skill.md file containing front matter (name, description, trigger conditions) plus any combination of supporting instructions, code snippets, resources, and nested role definitions. You zip it up, drop it into Claude’s new Capabilities settings tab, and Claude now treats the whole package as a system-wide plugin — triggering it automatically when relevant, without you having to re-paste your instructions into every new chat.

The live demo made the concept concrete. Lou had built an “AIM Writing Team” skill that chains six sub-roles: Researcher, Strategist, Writer, Editor, Publisher, and an Orchestrator that sequences them. Ask it to produce a 1,000-word article and it runs the full pipeline unprompted — each role scoring its own output, passing a summary to the next stage, and writing improvements to a persistent JSON memory file so the skill gets better with each run. Call a specific sub-role by name and the orchestrator routes to just that one.

One practical gotcha: you can only have one skill per zip file. Lou hit that wall and had to consolidate his original multi-file structure into a single orchestrated package before Claude would accept it.

The Real Question: Why Did This Go Viral?

Lou and Donald both had the same reaction — “is that it?” — and neither was being dismissive. The group’s working theory is that practitioners who’ve been building with Claude Projects, custom GPTs, and structured prompts for two years have essentially been doing skills all along. The new wrapper makes it system-wide rather than project-scoped, and it makes it portable enough to work (with some hacking) in ChatGPT and other frontends.

Hot take: Skills are less about new capability and more about making prompts first-class citizens of the Claude OS. The strategic move for Anthropic is ecosystem lock-in — give users a way to build deep, personalized tooling that lives inside Claude and doesn’t port cleanly elsewhere. It’s the App Store play.

Dirk’s CEO Experiment and the “AI Board” Concept

Before Lou joined, Dirk had shared something the group was still riffing on: he’d built a kick-ass AI CEO to give him direct, unfiltered strategic direction. Kasimir extended the idea by describing his own AI board — multiple advisor personas he consults collectively. When Lou demoed Skills, Dirk immediately connected the dots: “I think now that could be skills. I could create a skill board.” Lou agreed without hesitation: “You can call the board together anytime.”

This is one of the more generative use cases to emerge from the session — rather than one monolithic assistant, you build a panel of distinct roles (CEO, CFO, Devil’s Advocate, Industry Expert) each as a skill, with an orchestrator that convenes them when you need a strategic decision reviewed from multiple angles.

API vs. MCP: Lou’s Honest Field Report

Lou has been running an MCP server in a live legal AI app, and the verdict is mixed. The promise of MCP — a universal, portable interface that any frontend can query — is sound in theory. In practice, Lou found it slow and context-hungry. Switching to a direct API call on the same task was nearly instantaneous.

His current position: “I’ve used it, I know how it works, I learned it, but I’m not ready to deploy it across all my assets. We’re still in alpha stages of this framework.” He prefers API servers for speed.

Kasimir’s parallel point: Claude’s 200K context window sounds enormous until you’re in a long iterative coding or documentation session and you hit the limit at the exact moment you’re closest to done. His longer-term goal — routing the same conversation through different models based on what each does best — remains aspirational.

The Maintenance Cost You’re Not Calculating

The conversation took a useful turn into business fundamentals when Lou invoked two mentors. First, T. Harv Eker’s rule: add 30–40% annually on top of any purchase price to account for maintenance, downtime, integration work, and learning curve. Then Warren Buffett’s future-value framing: that $100,000 purchase, compounded at 20% over 10 years, is actually a $2 million decision.

Applied to AI tooling, the message was direct: “If I’m a solopreneur, the question is always — if I were to call a client to make an offer instead of spending my time on this, what’s going to be better for my business? And 90% of the time, it’s talk to the client.”

Mazie, drawing on her industrial engineering background at Johnson & Johnson, made the same point from the factory floor: “If the line was working and producing what it was supposed to produce, you didn’t touch it.” Kasimir backed it up with a Nokia story — a printer driver update for a label-making machine stopped an entire phone factory.

The framework: be an eager early adopter for experimentation, but a deliberate late adopter for deployment.

Community Corner

Bally kicked off with a story worth sitting with. She’d been introduced to a woman named Wendy — an English coach living in Thailand, serving authors and coaches who want AI-powered support. Wendy uses nothing but GPT-4o custom GPTs, refuses to engage with prompt engineering, and charges £600/month recurring to clients who meet with her weekly. Six months in, she has recurring clients and a portable business. Bally’s reaction: “She’s kicking ass — and what we’re doing here is so much more than that.”

The story landed as a grounding reminder. You don’t need the full stack to build a valuable, repeatable service. Someone simpler than you is already charging for it.

Lou closed with personal news. His father-in-law, diagnosed with stage 4 lung cancer and given days to live, has responded to an experimental immunotherapy trial in Thailand. The tumor is nearly gone. The drug costs $2,500/month — accessible there, unattainable most places.

Try This Week

Build your first Claude Skill as a personal prompt library entry. Pick one task you do repeatedly — a weekly report, a client proposal, a content outline — and package it as a skill. Write a skill.md file with front matter (name, description, trigger phrase) and your best prompt as the instruction body. Zip the folder, upload it in Claude’s Capabilities settings, then start a fresh chat and see if Claude activates it without being told to.

Time investment: 30 minutes. Then note whether Claude triggers it correctly on a follow-up session. That’s the real test.