AI Engineering from the Ground Up
The 5-rung spine as a beginner-friendly walkthrough. Prompt → skills → agents → loop → graph, with Karpathy autoresearch as the concrete case.
WORKBOOKS · THE 5-RUNG SPINEPROMPT → SKILLS → AGENTS → LOOPS → GRAPHS
CHRISTIAN.T.MACIONUTC+85 LAYERS2 WORKBOOKSOWNER-VERIFIED
Prompt → skills → agents → loops → graphs.
Modern AI systems are built in five compounding layers. Most teams plateau at the loop rung. The graph rung is the productivity unlock.
layers
5
the spine
workbooks
2
shareable
agents
11
in STELLA
pages
~62
PDF total
The spine
Five compounding layersn/a pick your rung.
Every layer is a force multiplier on the one below it. Most teams stop at the loop. The graph is where the agents start to share memory instead of re-deriving context.
- 01
Prompt
A single instruction to a model. Stateless, ~5 layers.
/ everyone - 02
Skill
A structured folder of instructions + scripts + refs. Loaded progressively.
/ Claude Code - 03
Agent
A long-running delegated task. Owns its context, tools, dispatch.
/ researcher - 04
Loop
An iterative cycle with eval-gate enforcement. The 5-must-have contract.
/ engineer - 05
Graph
Typed entities + relations + provenance. Shared memory across agents.
/ architect
Shareable
Two workbooks, one claim.
One big claim. the 5-rung spine. Two levels of audience. beginner, technical. Each one is a self-contained PDF + shareable HTML.
Graph Engineering for Everyone
The Königsberg-to-multi-agent arc. Why context windows fail, how to build a four-stage Extract · Resolve · Assemble · Query pipeline, where it fits in your stack.
The 53-agent office runs on the same 5-rung spine.
If you want to see how the spine is wired into a working multi-agent platform, the AI lane page is the live tour.