Microsoft's AI costs now exceed its payroll. You don't need their budget — you need the right workflow.
This is a build guide, not an essay. Four tools, four build types, the exact prompts to paste into Claude Code, the mistakes that waste a weekend, and what all of it actually costs. Everything below fits into two days and about twenty dollars of API credit.
Enterprise AI now costs more than the employees it replaced
Microsoft's internal AI spending exceeded its human workforce costs this year. Uber burned through their entire 2026 AI tools budget in just four months. A VP at Nvidia said it plainly: "The cost of AI for my team was more than humans."
These companies are bleeding money because they're using AI like a department — thousands of licences, enterprise SLAs, managed infrastructure.
Copilot licences at scale. Managed compute. Vendor lock-in. Costs that compound with every seat, every API call, every team that needs access.
Claude Code + Replit free tier + Supabase free tier. One weekend. One deployed, working product. No procurement. No DevOps team.
Four tools. Every AI product you can imagine.
No tool tourism. Each one earns its place by shipping real outcomes — used together, they replace an entire engineering team for prototype-to-MVP work.
| Tool | Layer | What it does |
|---|---|---|
| Claude Code | AI layer | Your AI co-developer. Writes, debugs, and iterates on code from plain-English instructions. Replaces hours of Googling. |
| Replit | IDE + deploy | Browser-based dev environment with one-click deploy. No local setup, no Docker, no server configuration. |
| Supabase | Database | Postgres database + auth in minutes. Free tier handles thousands of rows and concurrent users. |
| n8n | Automation | Visual automation builder. Trigger workflows, connect APIs, build agents — no code required. |
Pick your build. Copy the workflow.
Each build type below comes with the exact steps and Claude Code prompts I use. Pick what you're shipping — the workflow adapts.
Four shapes cover nearly everything worth building in a weekend. Step through them to see what each one costs you in time, what it runs on, and what you are left holding when it finishes.
AI Chatbot
A support bot that remembers the conversation. Six steps: create the Repl, add the key as a secret, scaffold the whole app with one prompt, debug by describing the failure, move storage to Supabase, deploy.
Deployed chatbot with conversation memory. Live URL. Shareable.
Automation
No code at all. A trigger, an HTTP request to Anthropic, and a destination node. Five steps from signing up to something that runs itself every morning.
A workflow running on schedule, 24/7, zero maintenance.
Full MVP
A real product: one-sentence spec, one scaffold prompt, an hour of plain-English iteration, then deploy. Four steps, and the last one is shipping before it feels ready.
A shareable, deployed product with real users by Sunday.
AI Agent
An agent takes action rather than answering. One job, one endpoint on Replit, one n8n trigger that hits it on a schedule, one destination.
An agent running 24/7 without you — real actions, real output.
Every step, prompt and rule from all four workflows is written out in full below. Nothing is hidden behind a tab.
AI Chatbot, deployed in 45 minutes
Forty-five minutes, three tools, one deployed chatbot that remembers the conversation.
AI Chatbot
Claude Code · Replit · Supabase. Deployed chatbot with conversation memory. Live URL. Shareable.
01 · Open Replit — create a Node.js Repl
3 min. Go to replit.com → New Repl → Node.js template. Name it. This is your full dev environment — no installs, no config.
02 · Add your Anthropic API key as a Secret
2 min. Tools → Secrets → Add ANTHROPIC_API_KEY. Never hardcode keys in files.
ANTHROPIC_API_KEY = sk-ant-api03-...
03 · Prompt Claude Code to scaffold the full app
5 min. Hit the AI button in the Replit sidebar. Paste this exact prompt.
Build a customer support chatbot web app. - Uses Anthropic API (key in env: ANTHROPIC_API_KEY) - Reads FAQ from a hardcoded JS array - Keeps conversation history in memory per session - Clean chat UI (white bubbles, dark bg) - Express.js backend, vanilla JS frontend Run on port 3000.
04 · Debug by describing the problem to Claude
10 min. When something breaks, don't Google it. Tell Claude what happened.
The chatbot replies but loses history after each message. Expected: context should persist across the full session. Error: [paste exact error here] Fix this.
Never say "it doesn't work." Always say what you expected, what happened instead, and paste the exact error. Claude fixes it in one shot.
05 · Connect Supabase for persistent storage
10 min. Create a free Supabase project → grab URL + anon key → add as Replit Secrets → ask Claude Code to upgrade.
Replace in-memory storage with Supabase. - SUPABASE_URL and SUPABASE_KEY already in env - Create table: conversations columns: id, session_id, role, content, created_at Store every message. Load full history on session start.
06 · Deploy — one click, live URL
2 min. Replit Deploy button → your app is live at a public URL. No DevOps. Share it. You shipped.
~$20 in API credits. Supabase free tier. Replit free tier. The same thing costs enterprises thousands per month.
The only hard part is step four, and it is a habit rather than a skill: describe the failure instead of naming it.
Automation that runs without you
Thirty minutes, no code, and a workflow that keeps running after you close the tab.
Automation
n8n Cloud · Claude API. A workflow running on schedule, 24/7, zero maintenance.
01 · Sign up for n8n Cloud (free tier)
3 min. n8n.io → Start free. Visual workflow builder — every node is drag and drop. No code required.
02 · Set your trigger
5 min. What starts the workflow? Schedule (runs every X hours), Webhook (URL hit), or Gmail (new email arrives).
Schedule trigger (8am daily) → fetch top AI news → Claude summarises in 3 bullets → sends to Telegram or Slack.
03 · Add HTTP Request node → Anthropic API
8 min. Add HTTP Request node. Configure as POST to Anthropic.
URL: https://api.anthropic.com/v1/messages
Method: POST
Headers:
x-api-key: YOUR_KEY
anthropic-version: 2023-06-01
Content-Type: application/json
Body:
{
"model": "claude-sonnet-4-20250514",
"max_tokens": 500,
"messages": [{"role":"user","content":"Summarise in 3 bullets: {{$json.content}}"}]
}04 · Route output to Slack, Telegram, or Sheets
5 min. Add final node. Map {{$json.content[0].text}} as your message body. That's the AI response.
05 · Activate — runs forever, for free
2 min. Execute once to test → if it passes, toggle Active. Your automation now runs on schedule with no maintenance.
The same thing consultants charge £50k to deliver. You did it in 30 minutes.
Every node above is drag and drop. The only text you write is the prompt inside the HTTP request body.
A full MVP in one weekend
One weekend, one spec sentence, one scaffold prompt, and an hour of iteration in plain English.
Full MVP
Claude Code · Replit · Supabase. A shareable, deployed product with real users by Sunday.
01 · Write a one-sentence spec before touching code
10 min. Fill in: "A web app that lets [WHO] [DO WHAT] so that [OUTCOME]." Nothing vague.
A web app that lets [founders] [paste a company URL and get an AI-written cold email] so that [they save 30 min per outreach target].
02 · Scaffold the full stack in one Claude Code prompt
15 min. One detailed prompt. Get the whole skeleton — backend, frontend, database, API calls.
Build a full-stack MVP: paste URL → AI-written cold email. Stack: Express.js + Vanilla JS + Anthropic + Supabase All keys already in Replit env. Pages: 1. Home — URL input + Generate button 2. Result — email output + Copy button System prompt for Claude: "You are a B2B sales expert. Write a cold email. Max 150 words. Subject line + body." Store: input_url, email_output, created_at in Supabase. Port 3000.
03 · Iterate in plain English — UI, logic, edge cases
60 min. Each round of changes = one Claude Code message. Be specific about what you want.
// Round 1 — UI Dark bg, white card result, mobile responsive. // Round 2 — UX Add loading spinner while email generates. // Round 3 — Edge case If URL scrape fails, show friendly error. Do not crash the app.
04 · Deploy and share before it's perfect
5 min. Replit Deploy → live URL. Share with 5 people before adding another feature. Real feedback beats more code.
Ship when 3 people can complete the core action without you explaining it. Not when it's perfect.
Notice how little of that weekend is spent typing code. Most of it is spent deciding what the thing is and then correcting it out loud.
An AI agent that takes action
Ninety minutes for something that takes action on its own: one job, one endpoint, one schedule.
AI Agent
Claude Code · Replit · n8n. An agent running 24/7 without you — real actions, real output.
01 · Define the one job the agent does
10 min. Agent = AI that takes action, not just answers. Decide: trigger → action → output. Keep it one job.
Every morning → scrape 10 Reddit posts from your niche → Claude picks top 3 → posts summary to your Slack.
02 · Build the action endpoint on Replit
20 min. Claude Code builds the API endpoint — receives data, calls the AI, returns a result. This is the agent's brain.
Build a POST endpoint at /agent/summarise.
Input: { "posts": ["post1...", "post2..."] }
System prompt for Claude:
"Pick the 3 most insightful posts for AI builders.
Return JSON: [{title, why_relevant, url}]"
Return Claude's parsed JSON.
Express.js. Key in env: ANTHROPIC_API_KEY. Port 3000.03 · Wire the trigger in n8n
20 min. Schedule trigger → RSS/API source → HTTP POST to your Replit endpoint → Slack send.
Schedule (8am daily)
→ RSS: reddit.com/r/MachineLearning.rss
→ Code node: extract top 10 titles + links
→ HTTP POST: your-repl.replit.app/agent/summarise
→ Slack: #ai-news → {{$json.content}}04 · Deploy as Always On — runs without you
5 min. Replit Deploy → endpoint stays live 24/7. n8n hits it on schedule. The agent runs forever without you touching it.
A 24/7 AI agent doing the job that would take a junior hire 30 minutes every morning. Automated, free, and fully yours to extend.
The endpoint is the brain and n8n is the alarm clock. Keep them separate and either one can be replaced without rebuilding the other.
Four mistakes that kill first builds
Most builders hit the same walls. Knowing where they are is half the fix. Open any row for what actually goes wrong.
What this means if you're building
If you're building in the AI space, the current landscape is a tale of two worlds: bloated enterprise suites vs. lean, modular stacks. Here's what actually determines who wins.
Speed is the only defensible moat
Enterprise AI is paralyzed by procurement, compliance, and multi-departmental sign-offs. If you are building lean, your advantage isn't just cost — it's the ability to ship, fail, and iterate while they are still drafting the project charter.Leverage > Feature Count
Don't try to build the next platform. Use tools like Claude Code and n8n as force multipliers. One person with the right workflow can output more functional code in a weekend than a 10-person enterprise team burdened by meetings and legacy debt.Modularity prevents vendor lock-in
Enterprise stacks are designed to trap you in an ecosystem. By building on portable infrastructure like Supabase and Replit, you remain agile. If a vendor changes their pricing, you have the flexibility to pivot without re-platforming your entire business.Shipping value beats building hype
Enterprise AI often focuses on building "AI features" that don't solve user problems. You are building solutions. When you ship lean, you aren't just saving money — you are forced to focus on the one thing that actually delivers value to the end user.Build with AI. Understand what you're building.
Deep technical breakdowns, build walkthroughs, and real AI product tutorials — for founders, operators, and developers who want to stay on the right side of the gap.
Microsoft's bill is not your bill. The stack above costs about twenty dollars and a weekend, and it ends with a URL you can send to someone.
Haroon K M · @ closefuture · haroon.pro. Share this piece with whoever told you AI was too expensive to build with.