Notes

Longer-form writing that doesn’t fit in a prompt or persona. New to all of this? Start with How LLMs Actually Work and read down.

Fundamentals

How LLMs Actually Work: next-token prediction, tokens, temperature, and why nothing persists
Context Windows in Practice: quality degrades by dilution long before the limit, and what hygiene helps
When to Trust Model Output: the reliability gradient from "transforming shown text" down to "citations from memory"
Reasoning Models and Thinking Tokens: what the working-out phase is, when it earns its cost, and what it doesn't promise

How Do I Give the Model My Data?: paste it, retrieve it, or train on it, in that order
Reliable Structured Output: from asking nicely to enforcing a schema, and designing schemas the model can fill honestly
Evals That Mean Something: real failure cases over synthetic ones, and when an LLM judge grades its own homework
Managing Token Costs: practical rules for keeping LLM API bills sane
Prompt Injection: why the model can't tell instructions from data, and how to cap the blast radius
Anatomy of a Persona: the Rubber Duck persona, line by line, and why every rule exists
Measuring Your Claude Code Spend: what the transcript data actually shows, and two habits that keep costs under control

Agents 101: an agent is a model in a loop with tools, and when you actually need one
Designing an Agent's Tools: bash vs dedicated tools, descriptions as triggers, and keeping secrets out of the loop
When to Use Sub-agents: what delegation buys you, when it backfires, and the cold-start problem
MCP in Practice: what the Model Context Protocol is for, when to skip it, and the gotchas nobody mentions
Working With a Coding Agent: plan-first, reviewable diffs, proof over claims, and knowing when to reset
Guardrails for Coding Agents: instructions vs enforced rules vs credentials, and matching the layer to the stakes

Running Local LLMs: when your own hardware beats an API, the Ollama-to-vLLM stack, and the memory math
How Big Models Become Small Ones: quantization, distillation, mixture of experts, and why parameter count stopped meaning much