LLM ✍️ Notes

Large Language Models (LLMs) — What to Remember

A memory‑friendly, hand‑annotated summary you can drop into any blog page. Mix & match the CSS classes below to get different “student highlight” looks.

Core Uses of LLMs

LLMs are general‑purpose text engines trained on large corpora. They excel at knowledge retrieval & synthesis, content generation, reasoning over structured/unstructured inputs, and language transformation (translate, rewrite, compress).

Common jobs: assistants, search & RAG, coding, data wrangling, agents, and automation.

Variant A — Classic Neon Highlighters

Retrieval‑Augmented Generation (RAG) pairs an LLM with a search index to ground answers in your docs, reducing hallucination. Add citations and chunking + embeddings to improve precision.

For analytics, prompt the model to produce JSON or tables and post‑check with validators.

Variant B — Gel‑Pen Underlines + Squiggles

Good prompts are goal‑tied, specify format, and constrain sources. Use few‑shot examples for style and system messages for rules.

Tip: keep output schemas stable; breaking schemas = brittle pipelines.

Variant C — Boxed Highlighter + Margin Notes

Production setups route queries by skill or cost. Cache frequent prompts and throttle long‑running tools.

  • Add guardrails (allowed tools, timeouts).
  • Log traces + prompts + model versions.
  • Evaluate with precision/recall, nDCG, human rubrics.

Variant D — Sticky Note + Callout

Remember: Grounding (RAG/tools) → Planning (agents) → Verification (checkers) → Feedback (logs/evals).

Use function calling to delegate to code, databases, or search; keep tool outputs deterministic.

Variant E — Mixed Markers (two‑stroke)

Safety requires policy prompts + classifiers and red‑team tests for jailbreaks. Mask PII, rate‑limit, and prefer server‑side secrets.

Deploy canaries: if the model leaks a token, alerts trigger instantly.

Memory Anchors (for quick recall)

Think G-P-V-F: Ground, Plan, Verify, Feedback. Pair each with examples you repeat in reviews.

Copy this file into your blog. You can apply the classes (.hl .yellow, .wavy, .box-hl, .sticky, .note[data-note], etc.) to your own text to get the same hand‑marked look.