How to Build Your Own AI: AI Agent Setup, Agent Automation and Build a Personal AI Assistant
If you want to build your own AI — a personal AI assistant that knows your business or create your own AI tools that run on your infrastructure — you need more than prompts. You need the infrastructure layer: agent memory systems, MCP server connections, and automation pipelines that hold context across sessions. These guides cover how to build that foundation from real builds, with specific steps and working code.
Running AI Locally on Windows
You do not need cloud subscriptions to run capable AI models on your own hardware. Two tools make local AI accessible on Windows: Ollama (terminal-first runner, supports Llama, Mistral, Qwen, and dozens more — free, no account required) and LM Studio (GUI-based, good for exploring and testing models without touching a terminal). Both are free to download and run entirely offline.
Recommended tools
- Ollama — terminal-first, scriptable, Claude Code-compatible
- LM Studio — GUI interface, good for first-timers
Minimum hardware
- 8 GB RAM — runs 7B models at usable speed
- 16 GB RAM — comfortable for 13B models
- GPU optional — CPU-only works, just slower
$ cd ~/how-to
How-To Wiki
The practitioner's guide to AI tools.
People are one prompt away from spinning up an AI agent but three years away from having a personal AI assistant that actually knows their business. The gap is AI infrastructure. These guides cover how to build that foundation from scratch: agent memory systems, MCP server connections, parallel agent pipelines, and security patterns that work without paid tools. Every guide is a working playbook from real builds — specific steps, real edge cases, no theory. Build your own AI content pipeline or automation stack and own it completely.
IDE Fundamentals
How to work with any AI IDE — Cursor, Windsurf, VS Code + Continue. Principles that apply everywhere: context windows, file references, rules, skills, composer vs inline.
CLI Tools
Claude Code standalone, Claude Code inside Cursor, OpenClaw, repo-level context engines. The "cron job" stability model.
MCP Servers
What MCPs are, how to add and manage them, stack-specific setups for GTM, content, and full-stack workflows.
Cost Efficiency
Model selection strategy, credit management, token budgeting. How to get maximum output without burning through your subscription.
Security
Fact-based security guidance: .gitignore, env vars, context engines, public vs private repo strategy. Debunking the fear.
Parallel Agents
How to run parallel agents, orchestration patterns, when to split vs sequence, real examples from the OS.
Comparisons
Head-to-head breakdowns of AI coding tools, workflows, and concepts. Practitioner perspective from someone who uses them all daily.
ABM Pipeline
Account-based marketing automation, personalized landing pages, analytics tracking, and the full ABM motion from signal to conversion.
Deployment Tools
Hosting platforms, agent deployment, SDK kits, and infrastructure tools for shipping agents and services beyond static sites.
Tool Evaluation
Independent evaluation frameworks for go-to-market tools, vendors, and agencies.
Which Claude model should you use?
If you are building with Claude Code or wiring up Claude to an agent pipeline, you will hit this question fast: Opus, Sonnet, or Haiku? Here is the short version.
| Model | Speed | Relative cost | Best for |
|---|---|---|---|
| Claude Opus | Slower | $$$ | Complex reasoning, hard coding problems, multi-step agents where quality is critical |
| Claude Sonnet | Fast | $$ | Daily coding, writing, agent tasks — near-Opus quality at ~5x lower cost. Default for most use cases |
| Claude Haiku | Fastest | $ | High-frequency tasks: classification, summarization, routing — anywhere you need scale or sub-second response |
Claude Opus vs Sonnet vs Haiku: when each makes sense
Opus is the right call when you are stuck on something genuinely hard — a complex refactor, a multi-file architectural change, or an agent that needs to reason through ambiguity without hand-holding. The quality jump is real. The cost jump is also real.
Sonnet handles 90% of what most builders do day-to-day. In Claude Code sessions it is the default for good reason — fast enough to feel responsive, capable enough to write production code without constant corrections.
Haiku is what you reach for when you are running something thousands of times: a nightly cron that classifies leads, a pipeline that routes content, a webhook handler that needs to reply in under a second. Not built for depth — built for throughput.
The practical rule: start with Sonnet. Upgrade to Opus only when Sonnet gets it wrong twice on the same problem. Use Haiku for anything you would not want to pay Sonnet prices to run at scale.
Related guides
Want to go deeper? See how to build your own AI companion free — a step-by-step log of setting up a personal AI assistant with no paid subscriptions.