Advanced curriculum · practical systems

Beyond the Chatbox

A technical curriculum for building persistent, stateful, and autonomous AI systems. Transition from stateless queries to high-efficiency agentic workflows.

Five modules, each a working technique rather than a definition. If you are new to any of this, start with Phase 1 and come back — this page assumes you already know what a token is.

Module 01

Procedural Memory: Agent Skills

Stop using monolithic system prompts. Agent Skills provide a modular framework for injecting procedural knowledge—how to perform specific tasks—without bloating the context window.

Progressive Disclosure

The agent doesn't load everything at once. It indexes metadata first, then loads the body only when a semantic match occurs.

Step through the three stages and watch the file on the right change:

path: /skills/pdf-handler/SKILL.md
---
name: pdf-handler
description: Extracts text and tables from PDFs, 
             fills forms, and merges files.
allowed-tools: [shell, python]
---

# Procedure
1. Analyze the PDF structure.
2. If encrypted, request user passphrase.
3. Utilize the Python `PyPDF2` library...
~120 Tokens
Module 02

The Memory Bank

LLMs are stateless. The Memory Bank pattern creates a persistent "external brain" using hierarchical Markdown files, reducing redundant explanations by 23%.

Click a file to explore its role

Understand how temporal state is orchestrated across sessions.

Project Root

📂 .cursor/memory/

File Content Focus

Select a file from the explorer

Module 03

Computational Economics

Agents resend their whole context on every step, so a long loop pays for the same tokens dozens of times. Caching the stable prefix is the single largest lever on that bill — and the one most teams never pull.

The Power of Prefix Caching

Prompt caching reduces latency by 85% and slashes input costs by 90% by storing Key-Value (KV) tensors. Success depends on placing static data at the absolute beginning of the prompt.

Standard Request Cost 100%
Cached Read Cost (Anthropic) 10%

Cache Architecture Comparison

Module 04

Compression & Model Strategy

Managing token exhaustion and choosing the right compute tier.

SLM

When to use SLMs?

  • • Routine document classification
  • • Tool call JSON generation
  • • Local RAG embeddings
  • • Edge/Mobile environments
∑

Rolling Summary

Uses a cheaper LLM to generate prose summaries of old logs.

LOSS RATIO: MEDIUM
⚓

Anchored Incremental

Maintains a structured JSON document, only updating with new facts.

LOSS RATIO: LOW
⚡

Kompact Proxy

Middleware that uses 8 adaptive transforms to strip syntax.

COST SAVING: 40-70%
Module 05

The Perfect Agent Lifecycle

Every module above, assembled into one loop. Read it as the order things happen in, not a list of features.

Initialize with cached Memory Bank files → Index metadata → Load required Skills → Run loop with incremental summarization → Externalize state to Memory Bank → End.