Agentic Basics

About

PH

co-host 01

Pingfan Hu

PhD Candidate, GWU. Heavy Claude Code user.

↗ pingfanhu.com

JH

co-host 02

Dr. John Helveston

Associate Professor, GWU. Open-source educator.

↗ jhelvy.com

Section 1
Assistants vs Agents

The opening problem

new model new framework 5 must-install tools 10× productivity SDK rewrite

▼ ▼ ▼

— the reframe —

Don't chase releases.
Build a workflow that works for you.

  "Everyone has their developing flow." — Karpathy
 

Two terms, one distinction

passive

Assistants

You speak. It answers.

ChatGPT Claude.ai Gemini

active

Agents

Runs shells. Edits files. Picks next.

Codex Claude Code Gemini CLI

Assistants answer. Agents act.

Agent = Model + Tools

today's most accepted definition

Agent

=

Model

+

Tools

Agentic engineering = giving the model the right tools to act.

Two shifts of focus

Two-year shift in where engineering effort goes

Prompt
Engineering

Did the model
hear what we said?

→

Context
Engineering

Does the model have
the right info?

  prompt ⊂ context  ·  containment, not replacement
 

When to use which

  ?   what's the task
 

one-shot

→ Assistant

· quick lookup
· polish a paragraph
· fits one chat

multi-step

→ Agent

· refactor across files
· test, fail, retry
· anything that loops

  workshop assumes agents
 

Section 2
Architecture

Two kinds of directories

directory 01

Home ~

· computer-level scope
· agent can touch anything
· global rules only

⚠ dangerous

directory 02

Project cd

· scoped to one project
· own CLAUDE.md + .claude/
· own conversation history

✓ where work lives

Quick install

~ — zsh

// install
$ brew install --cask claude-code
  ✓ installed
// enter project
$ cd ~/code/your-project
 
// launch
$ claude

  or run claude from VS Code · Positron · Kaku
 

The four components

Context

CTX · 01

What Claude reads.

CLAUDE.md rules/

Capabilities

CAP · 02

What Claude can act on.

MCPs plugins

Behaviors

BHV · 03

How Claude operates.

skills hooks

Agents

AGT · 04

Delegated execution.

subagents verifiers

Context — the foundation

CLAUDE.md

Project root. Keep short.

auto

~/.claude/CLAUDE.md

Every project.

global

.claude/rules/*.md

On-demand topical.

on demand

memory/

Auto-managed. Never synced.

ignore

Trigger to write a rule: "reproduce this reliably."

Capabilities — what Claude acts on

   $ claude
  

←──→

github gmail figma browser postgres slack

MCP

model context protocol

Servers Claude calls as tools.

Plugins

/plugin

Skills + MCPs + hooks bundled.

⚠ MCPs eat tokens. Use sparingly.

Behaviors — how Claude operates

Skills · slash commands

   ›
   /commit
   /review
   /init
   /think
   .md files
  

Hooks · shell on events

PostToolUse

Lint after edit

PreToolUse

Format on write

Stop

Build verify

  → Part 2: skill design
 

Agents — delegated execution

parent main Claude

research

read 12 files

refactor

isolated context

tests

run in parallel

verifiers

› test diff log › confirm "done"

A concrete project layout

your-project · tree -L 2

your-project/
├── CLAUDE.md // auto-loaded
├── .claude/
│   ├── rules/ // on-demand
│   ├── skills/ // slash commands
│   ├── agents/ // subagents
│   └── settings.json // hooks
└── docs/ai/ // shared docs

  shape varies — every piece has a home
 

Plan first, then act

single biggest behavioral lever

01

Plan mode

→

02

Write plan

→

03

Agree

→

04

Build

!

Sideways? STOP. Re-plan.

Section 3
Handy Tools

Don’t panic

the noise

10 must-install the only IDE switch to X 5 plugins, 10× framework Tuesday

the truth

No must-install kit.

Most of what you need is buildable.

Software stack

Obsidian

Markdown notes

Spokenly

Speech-to-text

P

Positron

VS Code IDE

Kaku

macOS terminal

  small · configurable · integrates via text or terminal
 

Extensions taxonomy

kind 01

Plugins

Marketplace bundles via /plugin

everything-claude-code claude-plugins-official

kind 02

MCPs

External servers as tools

Excalidraw Figma Gmail Calendar

kind 03

Skills

tw93's Waza

/think /design /hunt /check /learn

Custom shell commands

q-* · quick utils

$ q-brew // brew up
$ q-pull // pull repos
$ q-hide // hide windows
$ q-close // close apps

cd-* · shortcuts

$ cd-obsidian → ~/Vault
$ cd-website → ~/sites
$ cd-github → ~/code
 

  one line each in .zshrc  ·  real time saved daily
 

The actual hardest skill

"

Your success is determined by your ability to speak, your ability to write, and the quality of your ideas — in that order.

   — Patrick Winston, MIT
  

old

"Show me the code."

→

new

"Show me the prompt."

Three takeaways

✓

Assistants vs Agents

Agent = LLM + tools + loop.

✓

Architecture

Working dir = scope. 4 components. Plan first.

✓

Handy Tools

No must-install kit. Speak clearly.

Up next

— part 02 —

Skill Usage
& Design

ch 01

Common skills

ch 02

Design SKILL.md

ch 03

Share with symlinks