How to Run a Business on Claude Skills
Most businesses use Claude like a fancy search box. A few have installed it as an operational system — Business Brain, connected data, and 5–10 AI employees handling specific recurring work. Here's how that system actually works in production, and the twelve Skills every service business needs first.
Running a business on Claude Skills means configuring Anthropic's Claude Code as an operational system — a Business Brain (CLAUDE.md) capturing your context, connected data via MCP servers, and 5–10 AI employees (Skills) handling specific recurring work — rather than using Claude as a chatbot. It's the difference between having an AI tool and having an AI team.
This site runs on that system. 94 blog posts, complete marketing stack, CRM, email automation, analytics — single operator, no agency, no team. Everything you read here was written, reviewed, or edited by a Skill running on top of a Business Brain that knows how we work. The same system we install for clients powers the site itself.
What follows is the operator's version of this playbook — not "what are Claude Skills" (plenty of generalist guides exist), but what a Skill-run business actually looks like in practice, what breaks in production, and the specific 12-Skill inventory we've designed for service businesses based on running the full system on ourselves.
What "running on Skills" actually means
For most people, Claude is a chat window. You open it, paste context, describe what you want, edit the output, close the tab. Every session starts from zero. The leverage is modest — maybe 2–3× faster than doing the work yourself.
For a business running on Skills, Claude is a configured workspace. The assistant already knows your voice, your ideal client, your standards, your banned phrases, your usual processes. It has read access to your drive, your call transcripts, your past proposals. When you prompt it, it performs the work; when certain events happen (an email arrives, a deal closes, a client books), a Skill fires automatically. The leverage shifts from 2–3× to 10× plus reclaimed hours.
The distinction is architectural. A chatbot is reactive. A Skill-run business has committed context and configured capabilities — a system, not a tool.
The principle: generic AI produces generic output. AI loaded with your business context produces output that sounds like you made it. That's the entire game.
The three pillars of an AI operating system
Every Skill-run business we've seen or built has three layers. Each is necessary; together they compound.
1. The Business Brain (CLAUDE.md)
A single markdown file that captures how your business actually works: voice samples, ideal client description, service delivery standards, banned words, repeated workflows, things that signal a bad fit. Claude Code reads this before every task. The free Business Brain Builder generates a working draft in 10 minutes of conversation. Anthropic's official docs describe the CLAUDE.md convention in detail.
2. The Second Brain (connected data)
Your business knowledge, organized and queryable. This is where MCP servers (Model Context Protocol) matter — they're the connectors that let Claude read your Drive, your Notion, your email, your call transcripts. Without a Second Brain, every Skill still needs context dumped in by hand. With it, the system pulls what it needs automatically.
3. The AI Team (Skills)
5–10 task-specific Skills, each configured for a role in your business. A Skill is a reusable instruction set stored as markdown — Claude loads it and follows the instructions. Build once, run forever. More on the specific 12 we install below.
The 12-Skill inventory every service business needs
Here are the 12 Skills we've designed from running the full system on our own operations and from analyzing where service-business owners actually spend repeatable hours. Priority order reflects ROI — the first five typically capture 60–80% of the total time savings, so build those before any others.
| # | Skill | What it does | Typical hours/week reclaimed |
|---|---|---|---|
| 1 | Proposal Writer | Drafts client proposals from discovery notes using your voice and templates. | 3–6 |
| 2 | Follow-Up Manager | Writes follow-up emails for open threads, pending deals, and silent clients. | 2–4 |
| 3 | Meeting Scribe | Processes call transcripts into structured notes, action items, and follow-ups. | 2–3 |
| 4 | Content Writer | Drafts social posts, blog intros, and newsletter sections in your voice. | 3–5 |
| 5 | Client Communicator | Drafts routine client communications (status updates, check-ins, delivery notes). | 1–3 |
| 6 | Lead Researcher | Given a prospect name/company, compiles a briefing doc before calls. | 1–2 |
| 7 | Contract Drafter | Populates contract templates from deal parameters (needs legal review). | 0.5–1 |
| 8 | Onboarding Coordinator | Handles kickoff logistics — welcome emails, form collection, asset requests. | 1–2 |
| 9 | Report Generator | Produces weekly/monthly client reports from raw data in your format. | 2–4 |
| 10 | Inbox Triage | Sorts new email by intent (lead, client, vendor, noise) and drafts replies. | 2–3 |
| 11 | Analytics Digest | Summarizes analytics dashboards into plain-English weekly briefs. | 0.5–1 |
| 12 | Meeting Prep | Before each call, assembles prior-conversation context, open threads, talking points. | 1–2 |
Projected totals: running all 12 should reclaim 18–36 hours/week for a typical service-business owner. At a $150/hr opportunity cost, that's $2,700–$5,400/week in reclaimed capacity from a one-time install.
Note on the numbers: these ranges come from task-by-task time analysis across service-business workflows plus running the full system on our own operations. They're estimates, not reported averages. Your mileage depends on how repeatable your work is — highly bespoke consulting reclaims less; high-volume proposal work reclaims more.
How Skills compose into workflows
A single Skill doing a single task is useful. The multiplier comes from chaining them. Here's a workflow we've built and run on our own operations — the new-lead intake loop:
- Inbound email arrives. Inbox Triage classifies it as "new lead."
- Lead Researcher auto-fires, pulling public info on the prospect into a briefing doc.
- Meeting Prep drafts a pre-call document with relevant past-conversation context and suggested talking points.
- After the discovery call: Meeting Scribe processes the transcript into structured notes.
- Proposal Writer drafts a proposal from the notes. Owner reviews and edits.
- Follow-Up Manager is scheduled to check in if the deal goes silent.
Five Skills, one workflow, 90% of the lead-intake process automated. The owner spends time on the conversation and the judgment calls; Claude handles the structured work around it.
The 5-phase install framework
Here's the framework for the CAIO Build. It applies whether you're installing for yourself or hiring a done-for-you partner. The sequence matters — skipping phases produces Skills that break in production.
Phase 1 — Intake (30 minutes)
Capture the business's voice, clients, processes, tools, and standards via a structured conversation. The output is a draft CLAUDE.md. Our free Business Brain Builder runs this phase end-to-end in 10 minutes and produces a working file.
Phase 2 — Business Brain (1–2 days)
Refine the draft into a working CLAUDE.md. Add 2–3 real output examples (a proposal you'd send, a follow-up email you've written, a social post in your voice). Add banned patterns — the specific words, structures, or claims you never want in output. This is the single highest-leverage file in the entire system.
Phase 3 — Second Brain (3–5 days)
Wire up MCP servers for the data Claude needs to read: Google Drive, Notion, Gmail, call transcript service. For each source, audit access boundaries (you want read access, not write, on everything non-trivial). Organize scattered knowledge into structured locations Claude can find. This phase tends to take longer than planned; budget accordingly.
Phase 4 — AI Team (5–8 days)
Build 5–10 Skills from the inventory above, prioritizing the highest-frequency work. Each Skill gets: a role description, specific rules, at least one example of great output, a short quality checklist it runs before responding. Test each Skill on 3–5 real inputs before moving to the next.
Phase 5 — Handover (60 minutes + 30 days async)
Walkthrough of the complete system, recorded. Written documentation. 30 days of async support as the owner starts using the system in real work. This phase catches the edge cases that only show up in production.
DIY vs. Done-For-You installation
You can build this yourself. Many business owners do. But the tradeoff is real. Here's the honest comparison:
| Dimension | DIY | Done-For-You (The CAIO Build) |
|---|---|---|
| Out-of-pocket cost | ~$20/mo (Claude subscription) | From $5,000 one-time + optional $500/mo retainer |
| Time cost | 40–80 hours learning + 20–40 hours building + weeks of iteration | ~90 minutes of your time total (intake + discovery + handover) |
| Time to production-ready | 6–12 weeks typical (per our observation of DIY founders) | 2 weeks |
| Skills built in first pass | 1–3 (DIY typically stops after the first Skill works) | 5–10 complete Skills |
| Failure modes captured | Learned the hard way, post-launch | Built into the installation from prior installs |
| Ongoing maintenance | Your own ongoing work | Optional retainer ($500/mo) or handle yourself |
| Ownership | Full | Full — no lock-in, no SaaS dependency |
The math: if your effective hourly rate is $150+, the break-even on hiring someone to install it is fewer than 35 hours of your time. The break-even math assumes the install lands on the first try — in practice first Skills often need rebuilding. Our own first version did.
What breaks at scale (and how to prevent it)
Three failure modes we've identified from running the system ourselves, ordered by how likely each is to bite once it's in production:
1. Business Brain drift
The CLAUDE.md you wrote in month one stops matching how you actually work by month four. New offers, new client types, new standards. Output starts feeling slightly off. Fix: schedule quarterly reviews. 30 minutes is usually enough.
2. Skill sprawl without documentation
You keep adding Skills until you have 15 and can't remember which does what. Fix: name Skills with explicit verb-object patterns (proposal-writer-new-client, not proposal2). Maintain a Skills README. Consolidate overlapping Skills quarterly.
3. MCP server outages or permission changes
Google Drive or Notion revokes access after an OAuth refresh. A Skill that worked yesterday fails silently today. Fix: health checks that verify each MCP connection is live; alert when one fails. Our retainer includes this.
What Claude Skills can't do yet
Operator honesty matters more than marketing honesty. Here are the current limits. If any of these are dealbreakers, a Skill-run business isn't the right fit:
- Skills aren't native schedulers. They respond to prompts or hooks, not clocks. For scheduled runs (daily at 8am, weekly Mondays), you need external orchestration: cron, Make.com, GitHub Actions, or similar.
- Skills don't handle real-time voice. Phone calls, live video — Claude can process transcripts after, but isn't the right primitive for live interaction.
- Regulated client-facing work needs a human in the loop. Legal advice, medical recommendations, financial advice — Skills can assist drafting, never autonomously deliver.
- Complex integrations require custom work. Not every SaaS has a polished MCP server. For niche CRMs or industry tools, expect custom scripting.
- Skills are deterministic about instructions, probabilistic about outputs. They follow your rules, but any given output could be slightly different. Quality-check the first 10–20 runs before trusting any Skill in production.
What a week looks like on the system
Here's the workflow pattern once the system is running — drawn from how we operate the CAIO on its own install. Specific numbers describe the pattern we see in practice, not a single reported week.
Monday morning. Inbox Triage sorts weekend emails by intent — leads, vendors, noise. New leads already have briefings assembled by Lead Researcher. Review time per lead drops from ~30 minutes to ~5.
Tuesday. Discovery calls. Meeting Prep assembles each pre-call doc before the call. After each, Meeting Scribe processes the transcript; Proposal Writer produces a first draft. Editing takes ~20 minutes — the draft is typically 80–90% there.
Wednesday. Content Writer drafts social posts and blog outlines. Claude never nails the exact cadence on first draft, so the intro gets rewritten by hand — but the body and structure hold up.
Thursday. Client Communicator handles standard delivery emails. Follow-Up Manager nudges silent threads with context-appropriate check-ins, queued for approval before they send.
Friday. Report Generator produces weekly reports. Analytics Digest summarizes Search Console data into a one-page brief instead of raw dashboards.
The net shift: repeatable work that used to take 10–14 hours/week compresses to 2–4 hours of review. Roughly a full working day reclaimed every week — what you do with it is the point.
How to get started
Three paths, depending on where you are:
Stop figuring it out. Get it installed.
Most business owners don't need another course or tutorial. They need the system built for them. Pick the path:
- Start free: Build your Business Brain in 10 minutes with Claude. You leave with a working CLAUDE.md.
- Decide what to build: The AI Audit — $997, personalized report in 48 hours. Tells you the first 5–10 Skills your specific business needs.
- Skip the building: The CAIO Build — we install the complete system in two weeks. From $5,000. You own everything.
FAQ
What does it mean to run a business on Claude Skills?
Running a business on Claude Skills means configuring Anthropic's Claude Code as an operational system — a Business Brain (CLAUDE.md) capturing your context, connected data via MCP servers, and 5–10 AI employees (Skills) handling specific recurring work. It's the difference between having an AI tool and having an AI team.
How long does it take to install a Claude Skills operating system?
Professional installation takes about two weeks: intake (30 minutes), discovery call (60 minutes), two weeks of build work, and a 60-minute handover. Learning and building this yourself typically takes 40–80 hours of focused time plus several weeks of iteration before it runs reliably.
How many Claude Skills does a service business actually need?
Most service businesses need 5–10 Skills to cover the majority of repeatable work. Core Skills include Proposal Writer, Follow-Up Manager, Content Writer, Meeting Scribe, and Client Communicator. Additional Skills depend on the specific business — lead research, contract drafting, onboarding, report generation, analytics digests.
What's the difference between a Claude Skill and a sub-agent?
A Skill is a reusable instruction set saved as a markdown file — Claude loads it and follows the instructions. A sub-agent is an autonomous Claude instance that runs in parallel, typically for complex research or multi-step tasks. Skills are the right primitive for recurring operational work; sub-agents are for tasks requiring parallelism or isolation.
Do Claude Skills replace employees?
No — they replace repetitive tasks, not people. Skills handle structured, high-volume work (proposal drafts, follow-up emails, content first drafts, meeting summaries). People still handle judgment calls, client relationships, strategy, and anything requiring real-world context Claude can't access.
What does a Claude Skills installation cost?
The CAIO Build starts at $5,000 for complete installation (Business Brain + Second Brain + 5–10 AI employees, delivered in two weeks). Optional retainer is $500/month for ongoing employee additions, data refresh, and health checks. DIY costs are lower in dollars but typically 40–80 hours of time plus iteration cycles.
What can Claude Skills NOT do?
Claude Skills don't replace human judgment, don't handle real-time phone calls, can't autonomously execute regulated workflows (medical, legal advice), and need external scheduling (cron, Make.com) for time-based triggers. Skills aren't native schedulers — they respond to prompts or hooks, not clocks.
Related guides
- What Is Claude Code? The AI Agent That Actually Executes — start here if you're new to Claude Code itself.
- Claude Code Skills: How They Work — the Skills feature explained.
- CLAUDE.md File Guide — deep dive on the Business Brain file.
- MCP Servers Explained — how Claude connects to your data.
- What Is a Chief AI Officer? — the role we install.