LLM Tutorials and Notes
16 posts on llm, newest first.
This page collects every GeekCoding101 post tagged llm: hands-on tutorials on working with large language models, from Claude Code and MCP servers to AI agents you build yourself in Python. Newer posts are step-by-step builds and tests you can rerun on your own machine. Earlier posts are short explainers on foundation models, fine-tuning, quantization and model size.
Start here: pick a post by goal
Building agents and MCP servers
If you are new to agents, start with the agent post. It explains the agent loop and builds a working Python agent with tools on the Claude API. Next, read the MCP server post to expose tools, resources and prompts through the Model Context Protocol. Then read the MCP authorization post to add OAuth 2.1, so the server rejects requests without a valid token.
Running Claude Code and agent frameworks safely
Two posts focus on what an agent is allowed to do on your machine:
- The
--dangerously-skip-permissionspost covers what the flag does, why it refuses to run as root, and the Claude Code permission modes. It also covers safer options: allow rules, sandbox and auto mode. - The dsh-superagent deep dive covers installing DeepSeek’s dsh-superagent on a Mac and letting it fix a real GitHub issue. It also walks through the framework’s two-layer authorization gate.
LLM fundamentals
The 2024 posts are short reads. They explain the ideas the hands-on posts build on:
- foundation models vs task models
- fine-tuning a pre-trained model
- how parameter count affects inference speed
- quantization from FP32 to INT8
Serving models
The Ray Serve guide covers deployments, autoscaling, batching and multi-model pipelines. It also explains when to choose Ray Serve over FastAPI or Ollama.
- 8 min read
Claude Plans Compared: Pricing, Claude Code and API Costs
Claude plans compared: what Free, Pro, Max, Ultra, Team and Enterprise cost, which include Claude Code, and when a subscription beats pay-as-you-go API billing.
- 8 min read
GPT-6.2 Sol vs Astra vs 6.1 Sol: Tested, With Pricing
GPT-6.2 Sol vs Astra vs GPT-6.1 Sol (and Opus 5.5) on the same 9 graded API tasks: accuracy, speed, GPT-6.2 Sol pricing and the real cost per task.
- 24 min read
DeepSeek Harness Hands-On: What dsh Really Runs (and Sends)
A hands-on test of DeepSeek Harness (dsh 0.2.2): install it, count its 299 plugins, write your own tool, and see what each request sends to DeepSeek.
- 6 min read
Funplay MCP: Connect Claude Code to the Godot Editor (Hands-On)
Funplay MCP is an open-source MCP server for game editors. Install the Godot addon, start the server, connect Claude Code and run a first safe request.
- 8 min read
Claude Fable 5.2 Effort Levels, Tested: 26 Runs vs Opus 5.5 and Fable 5.1
Anthropic says Fable 5.2 at High matches Opus 5.5 for half the cost; an independent test found it 4.5x pricier. I ran 26 runs on three models at medium, high and xhigh, with costs from transcripts.
- 9 min read
MCP Authorization: Protect an MCP Server with OAuth 2.1
How MCP authorization works: the 401, protected resource metadata and OAuth 2.1 flow, plus a runnable Python SDK server that rejects requests without a valid token.
- 11 min read
Claude --dangerously-skip-permissions: What It Does and When to Use It
What claude --dangerously-skip-permissions really does, why it refuses to run as root, the six Claude Code permission modes, and safer ways to stop the prompts: allow rules, sandbox, auto mode.
- 27 min read
I Ran DeepSeek's dsh-superagent End to End: A Hands-On Deep Dive
A hands-on test of DeepSeek's dsh-superagent: I installed it on a Mac, had it fix a real GitHub issue end to end, and explain its two-layer authorization gate, Mandates and risks.
- 9 min read
What Is an AI Agent? How to Build One from Scratch in Python
What makes an AI agent different from a chatbot, the agent loop explained, and a working Python agent with tools built step by step on the Claude API.
- 8 min read
What Is an MCP Server? How to Build One in Python
What the Model Context Protocol is, how MCP servers expose tools, resources and prompts, and how to build, test and connect one with the Python SDK 2.x.
- 9 min read
Ray Serve Guide: Autoscaling, Batching, FastAPI & Ollama
What Ray Serve is, how deployments, autoscaling, batching and model composition work, with code, and when to pick Ray Serve over FastAPI or Ollama.
- 8 min read
Model Quantization: FP32 to INT8, PTQ vs QAT
What model quantization is, how FP32 weights map to INT8 with a scale and zero point, PTQ vs QAT vs dynamic quantization, memory math and trade-offs.
- 7 min read
Foundation Models vs Task Models Explained
What foundation models and task models are, how they differ in data, training and scale, and when to prompt, fine-tune or distill for a specific task.
- 6 min read
Fine-Tuning vs Pre-Training: How LLMs Become Specialists
What fine-tuning is, how it differs from pre-training, the steps to fine-tune a pre-trained model, LoRA vs full fine-tuning, and the risks to watch.
- 6 min read
Parameters vs Inference Speed on Phones
Why phone AI models are smaller than GPT-4: how parameters, memory and hardware set inference speed, and how quantization, pruning and distillation help.
- 6 min read
Find and Remove Duplicate Photos with a Python Script
A Python script, written with ChatGPT, that finds duplicate photos by xxhash checksum and cleans them up, with cached checksums for fast reruns.
FAQ
Where should I start if I’m new to LLMs? Read the foundation models post, then the fine-tuning post. Together they cover the difference between a general pre-trained model and one trained further for a specific task. After that, the AI agent tutorial is a good first build.
What is the difference between an AI agent and a chatbot? A chatbot answers one message at a time. An agent runs in a loop: the model decides to call a tool, sees the result, and keeps going until the task is done. The AI agent post builds this loop step by step in Python.
What is an MCP server? It is a program that uses the Model Context Protocol to offer tools, resources and prompts to an LLM client. The MCP server post shows how to build, test and connect one with the Python SDK.
Is claude --dangerously-skip-permissions safe to use?
It removes Claude Code’s approval prompts, so the agent acts without asking first. The dedicated post explains when that is reasonable. It also covers safer ways to cut down on prompts, such as allow rules or a sandbox.
Are these posts free? Yes. Every post on GeekCoding101 is free to read, with no sign-up.