By Jereme Strange · also on Strange Intel
You probably landed here from one video. There are a lot more. This is the full library of every Claude-Code-for-business setup we've shared — real steps, copy-paste commands, no fluff. Bookmark it.
Jump to: The Full Stack · Memory · The Ops Layer · The Control Layer · 3 MCPs for Small Business · Start a Side Hustle with Claude · Turn Claude into an AI Worker · Put Your Business on Autopilot · Let Claude Design Your Website · Plan a Month of Content in One Prompt · The 3 MCPs That Matter Most · The 3 APIs That Run a Business · A Real Day, Run by Voice · Voice-to-Execution: Set It Up by Voice · Research That Actually Knows Your Business · Run Claude Headless · Four Problems, One Morning · Anthropic's Free Courses · Turn Claude Into Your Business Coach · Get Found by AI · A Second Claude on Its Own Telegram · Keep Claude Aligned · My Actual Daily AI Stack
The five layers behind an AI that runs a business, not a chatbot.
1. A rules file (CLAUDE.md) in your project root, with sections ## Brand Voice, ## Hard Rules, ## Pricing — it reads it every session. Run /init and Claude drafts it for you.
2. Persistent memory — a /memory folder, one fact per file.
3. Real hands: claude mcp add playwright -- npx @playwright/mcp@latest (a real browser); add more for desktop + socials.
4. API wiring — a free CLI like gog for Gmail/Calendar, the official MCP for your store.
5. A SessionStart hook in ~/.claude/settings.json so it's briefed before you type. (exact snippet in the Memory section below).
Most AI starts every session from zero. This is how ours comes online already knowing the business.
1. Make a /memory folder. One .md file per fact, each starting with frontmatter (name, type), then the fact. Claude pulls them on its own.
2. Add persistent recall: pip install nmem, then: claude mcp add neural-memory -- nmem-mcp — our own open-source Adam Framework. It remembers across every session.
3. Auto-load at startup: a SessionStart hook in ~/.claude/settings.json that compiles your memory before you type. The exact snippet:
{"hooks": {"SessionStart": [{"hooks": [{"type": "command", "command": "python C:/your-vault/compile_briefing.py"}]}]}} — swap the command for whatever builds your briefing (even just: claude -p "read memory + lanes + followups, write BRIEFING.md").
The layer that turns Claude Code into an employee that runs your day.
1. Make two files: LANES.md (one heading per active project) and FOLLOWUPS.md (a table — who / thread / next action / age).
2. Tell Claude to sweep them morning, afternoon & end-of-day and report only what moved. That's your automatic standup.
3. Schedule it: a Task Scheduler (Windows) or cron (Mac) job runs a headless command on a timer — e.g. claude -p "check my inbox — flag only what needs me" every 4 hours. It only pings you when something's real.
1. Install the bridge: /plugin install telegram → /telegram:configure <your BotFather token>.
2. Pair your phone: /telegram:access pair, send the code — now you voice-note it from anywhere.
3. Voice replies: pipe answers through edge-tts (free) in a voice you pick.
4. Routines: a scheduled task runs Claude on a timer — 6am email triage, etc.
5. Lock it down: an allow-list in settings.json so it acts without asking, with walls around anything risky. The exact snippet:
{"permissions": {"allow": ["Read", "Edit", "Bash(gog:*)", "Bash(git:*)"], "deny": ["Bash(rm:*)"]}} — allow what it may do freely, deny what it must never touch.
1. claude mcp add --transport http stripe https://mcp.stripe.com → "show this month's revenue / send that invoice".
2. claude mcp add --transport http hubspot https://mcp.hubspot.com/anthropic (free CRM) → "add this lead and draft a follow-up".
3. claude mcp add --transport http canva https://mcp.canva.com/mcp → "make a promo graphic for my June sale".
First run: type /mcp inside Claude Code to sign in to each one (OAuth — no API keys to handle).
1. Zapier — grab your personal server URL at zapier.com/mcp, then: claude mcp add --transport http zapier <your-url> → describe an automation, it builds the Zap; resell it as a monthly service.
2. claude mcp add --transport http apify https://mcp.apify.com → "scrape every dentist in Miami with email"; sell the list.
3. Drop in an ElevenLabs API key → Claude narrates faceless videos on demand.
1. Make a skill: a folder in ~/.claude/skills/ with a SKILL.md (name, description, steps) — now /your-skill runs the whole workflow.
2. Use /goal "close 5 leads by Friday" — it loops on its own until the condition's met.
3. Hit Shift+Tab for Plan Mode so it shows the plan before touching anything.
1. Google Workspace: a free CLI like gog → "check my inbox and draft replies for approval".
2. Drop an ElevenLabs API key in your env → voiceovers on command.
3. claude mcp add --transport http apify https://mcp.apify.com → scheduled lead-scraping that drops into a sheet.
1. /plugin install frontend-design
2. Tell Claude: "design a landing page for my business" (name your services + vibe).
3. It builds the whole site — hero, services, pricing — and opens the preview. Tweak in plain English.
1. Calendar hands: a free CLI like gog (github.com/openclaw/gogcli) wires up Google Calendar — Claude books every post onto it with real commands.
2. claude mcp add --transport http hf https://huggingface.co/mcp (free AI images)
3. Say: "plan my June content and add it to my calendar." It writes every post — hook, caption, hashtags, even a thumbnail — onto your calendar.
If you only add three things, add these. They turn Claude from a chat window into something with a browser, your machine, and a memory.
1. claude mcp add playwright -- npx @playwright/mcp@latest — gives Claude a real browser. It browses, clicks, fills forms, and researches like you would.
2. claude mcp add desktop-commander -- npx -y @wonderwhy-er/desktop-commander — gives Claude your machine: files, apps, terminal. It works your computer for you.
3. pip install nmem, then: claude mcp add neural-memory -- nmem-mcp — our own open-source Adam Framework. Memory that survives every session.
That's the base layer everything else sits on — eyes, hands, and a memory.
MCPs give Claude tools. APIs give it your actual accounts — your store, your inbox, your leads — so it runs the business live, not a dashboard.
1. Shopify — the honest version: there's no magic one-liner for your own store's orders. In Shopify admin: Settings → Apps and sales channels → Develop apps → Create app → give it read_orders + read_products scopes → install → copy the Admin API access token. Hand Claude that token (env var) + your .myshopify.com address and it queries the Admin API directly — "how many orders came in today?" now runs against your real store. ~10 minutes, once. That's exactly how we wired it for a client.
2. Google Workspace — a free CLI like gog wires up Gmail, Calendar, and Drive: say "add a client call tomorrow at 2pm" and it's on the books.
3. claude mcp add --transport http apify https://mcp.apify.com → "pull fresh leads from Google Maps" straight into a sheet, on a schedule.
Three connections and your AI is reading real orders, booking real calls, and pulling real leads.
Everything in that video ran through one setup. Here's exactly how to build it so your whole day moves by voice note.
1. The bridge: in Claude Code run /plugin install telegram → /telegram:configure with a free bot token from @BotFather → /telegram:access pair. Now voice notes from your phone become instructions.
2. Voice back: pip install edge-tts — it answers you out loud, in any voice you pick.
3. The watchdogs: a scheduled task runs Claude on a timer (bot checks, email triage) and only pings you when something's real.
4. The walls: a permissions allow-list in ~/.claude/settings.json so it acts without asking — inside the limits you set. (Exact snippet in the Control Layer section above.)
That's the full loop: you talk, it executes, it reports back. A real day, hands-free.
This is the one from the use-cases video — voice-note "set up an auto-DM on my new post" and it does the rest. Here's the exact setup.
1. The voice bridge: in Claude Code run /plugin install telegram → /telegram:configure with a free bot token from @BotFather → /telegram:access pair. Now you text or voice-note it from your phone.
2. Give it hands: launch Chrome in debug mode — chrome --remote-debugging-port=9222. Now Claude drives your real browser with Playwright/CDP, clicking through setups like the auto-DM exactly like you would.
3. Voice back: pip install edge-tts — it replies out loud in any voice you pick.
4. Scope it: a permissions allow-list in ~/.claude/settings.json so it acts without asking, inside the limits you set. (Exact snippet in the Control Layer section above.)
Now you just voice-note the task and it runs the clicks for you — that's voice-to-execution.
Generic AI research gives you generic answers. This is how ours comes back sharp — specific to your business, verified, and cited.
1. Make it an agent: drop a research agent file in ~/.claude/agents/ with the rules baked in — "verify across multiple sources, give exact numbers not vague ranges, free options before paid, and flag anything you couldn't confirm." The whole file can be this small:
---
name: deep-researcher
description: research with verification rules baked in
tools: WebSearch, WebFetch, Read, Write
---
Verify across multiple sources. Exact numbers, not vague ranges. Free options before paid. Flag anything you couldn't confirm.
2. Give it YOUR context: a CLAUDE.md or memory file with your business, who you serve, and your constraints — so every research question goes out razor-specific to you, not generic.
3. Isolate it: tell Claude to run research in a subagent — it fires 20+ searches and writes the findings to a file, without cluttering your main chat.
4. One command: voice-note or type the topic — it plans, searches, verifies across sources, and hands you back a cited brief.
That's the difference between a search engine and a researcher who already knows your business.
No chat window, no babysitting. Headless mode turns Claude into a process you can pipe anything into — and that other scripts can call.
1. One-shot it: claude -p "your instruction" — it runs the full agent loop (tools included) and exits. Pipe anything in: cat error.log | claude -p "what broke?"
2. Give it memory: add --continue so every call resumes the last session and remembers.
3. Give it a job: --append-system-prompt "you run ops for this store" — its identity survives every call.
4. Scope its hands: --allowedTools "Read,Write,Bash" — it acts without asking, inside walls you set.
5. Make it an employee: a small bot script pipes Telegram messages into that exact command on any spare machine — ours writes its own handoff doc and memory files every night.
That's an AI worker running on a spare laptop, no window open, doing the job while you sleep.
The real trick from that video isn't the tools — it's that you don't hunt for them. You tell Claude the problem and it finds, installs, and wires the vetted option. Here are the four it pulled:
1. gog — one CLI for Gmail, Calendar, Drive & Sheets (github.com/openclaw/gogcli). Claude reads your email and calendar with real commands.
2. SumatraPDF — silent printing from one line: SumatraPDF -print-to "Your Printer" -silent file.pdf. No windows, no clicks.
3. faster-whisper — pip install faster-whisper — voice-note transcription that's fast and stops butchering names.
4. VoiceFixer — pip install voicefixer — free, studio-grade speech restoration on your own machine.
The answer usually already exists. Tell Claude the problem and let it find the tool — that's the whole skill.
Want to actually learn this from the company that makes Claude? It's all free, self-paced, and every course gives you a real certificate for LinkedIn.
1. Go to anthropic.skilljar.com — sign up with any email (no Anthropic account needed). All free, all self-paced, certificate on every one.
2. Start here: Claude 101 (everyday use) → AI Fluency: Framework & Foundations (the thinking framework) → Claude Code 101 (when you're ready to build).
3. Going deeper: Building with the Claude API, Introduction to Model Context Protocol, Intro to Claude Cowork — same catalog, same price: free.
I've completed ten of them — they're on my resume and my LinkedIn. If you'd rather skip the learning curve and just have this stuff already running your business, that's what we do.
Most coaching forgets your context between sessions. This one holds your full picture, every day, and tells you the truth.
1. Give it the role: drop your goals, your numbers, and "coach me — hold me to this" into a CLAUDE.md / project instructions file. Now it coaches with YOUR full context, zero judgment.
2. Make it daily: one scheduled check-in — "what moved yesterday, what's today's one needle-mover?" Accountability on a timer.
3. Make it honest: tell it to challenge you with data, not cheer for you. Ask "what am I avoiding?" once a week and brace yourself.
A coach that knows every number, never forgets, and won't let you off the hook — running for free on your own machine.
Search changed: the AI picks the business now, and it reads the code under your site, not the pretty front. The files that get you surfaced:
1. robots.txt — explicitly allow the AI crawlers (GPTBot, ClaudeBot, PerplexityBot, Google-Extended). Not allowed = invisible.
2. Schema (JSON-LD) — LocalBusiness + Service + FAQ, so the AI knows who, what, and where you are.
3. llms.txt — a plain-text summary AI can ingest. Almost nobody has it = easy edge.
4. Consistent NAP, direct-answer copy, clean fast HTML, and a sitemap.
Full breakdown with the exact code: Get Found by AI
Run a second Claude on its own Telegram — same subscription you already pay for, zero extra cost. Mine's my kid's sidekick: its own bot, its own rules, locked down safe.
1. Make it its own vault folder and drop a CLAUDE.md inside — that's its rulebook (who it is + what it's allowed to do); run /init and Claude drafts it for you.
2. Give it its own bot: /plugin install telegram → /telegram:configure with a SEPARATE token from @BotFather (new token = its own bot, fully separate from yours).
3. Pair the phone that'll use it: /telegram:access pair → send the code.
4. Lock it down: an allow-list in settings.json so it only does what you've approved (exact snippet in the Control Layer section). Now it lives on its own Telegram and anyone you've paired can message it like a person.
Built-in memory helps, but it's limited — it drops old context and won't keep your whole business straight on its own. The fix isn't more memory; it's a few files that point to each other, plus one rule that keeps them current.
1. Four things in your project root: MEMORY.md (durable facts — timezone, key clients), LANES.md (what you're working on), FOLLOWUPS.md (who owes you what + the next action), and a /trackers folder (one file per client or project = your source of truth).
2. Make MEMORY.md the hub: it holds the permanent facts AND points to the live files.
3. One rule in CLAUDE.md: "At session start, read memory + lanes + followups + trackers; log new facts to the right file; keep them current." It re-reads them every session and updates as things change.
4. A SessionStart hook in ~/.claude/settings.json compiles all of it into a morning briefing (exact snippet in the Memory section) — you wake up to meetings + what's due instead of a blank slate.
Three layers — and here's the exact setup, not just the names:
1. THE VOICE LOOP (run it from your phone): /plugin install telegram → /telegram:configure with a free token from @BotFather → /telegram:access pair. Incoming voice notes: pip install faster-whisper transcribes them locally. Voice back: pip install edge-tts — it answers out loud. Voice in, voice out, from anywhere.
2. THE HANDS: launch a real logged-in Chrome — chrome --remote-debugging-port=9222 — and Claude drives it like a person (bypasses bot detection). Plus a free CLI like gog (github.com/openclaw/gogcli) for Gmail, Calendar & Drive from one command line.
3. THE BRAIN: memory files + LANES.md + FOLLOWUPS.md + a /trackers folder, all read at the start of every session — that's the ALIGNED setup one section up, snippets and all.
That's the whole stack: you talk, it acts in the real world, and it remembers your business.
Claude Code built us a mission-control dashboard for itself — a local web app (Node + Express on localhost) with five columns: a real terminal, the live task board, what Claude is doing right now, follow-ups, and a live mirror of our Telegram conversation. The trick that makes it honest: 1. The data source is Claude Code's own session transcript (~/.claude/projects/<your-project>/*.jsonl) — the app tails that file and renders it live, tool calls on the left, messages on the right. It can't show fake status because it only renders what actually happened. 2. The terminal panel is real (node-pty + xterm.js) — launch claude inside it and the dashboard IS your terminal; a mic button pipes local Whisper transcription straight into the same session. 3. To get yours, tell Claude Code: "build me a local dashboard that tails my session .jsonl and shows my task list, tool calls, and messages live." Ours took one day. Keep it bound to 127.0.0.1 — the terminal endpoint is a real shell, never expose the port.
That's the whole point of SAM. We install this entire system on your machine, seeded with your business, remotely — in days, not the months it took us to figure out. You own everything: the AI, the data, no lock-in.
Want this in your business? Free 30-minute consultation