SCHEDULE

Week 1 Install and Claude Basics
Week 2 Research and Writing with Claude
Week 3 Skills, MCPs, and extending open source projects
Week 4 Build your own software part 1
Week 5 Build your own software part 2
Week 6 Build your own software part 3
Week 7 Demo Day

TODAY

1. How text generation works
2. Writing with AI
3. Research agents
4. Hands-on exercises

TOKENS

Text is broken into tokens — roughly word fragments.

“unbelievable” → [“un”, “believ”, “able”]

The model predicts the most probable next token given all previous tokens.

BUT IT’S MORE THAN THAT

“Next token prediction” is the training objective, but modern models go beyond this in practice.

BUT IT’S MORE THAN THAT

• Multi-token prediction: newer models are trained to predict several tokens ahead at once, not just the immediate next one
• Speculative decoding: models draft multiple tokens in parallel, then verify them
• Extended thinking: models like Claude and o3 “think” internally before producing output, effectively planning across many tokens

BUT IT’S MORE THAN THAT

The core architecture is still autoregressive, but the behavior increasingly looks like reasoning in chunks rather than word-by-word.

CONTEXT WINDOW

The context window is everything the model can “see” at once.

• GPT-5: ~400K tokens
• Claude: ~200K–1M tokens
• Gemini: ~1M+ tokens

Everything you send (system prompt, conversation history, files) counts against this limit.

CONTEXT WINDOW

2 weeks ago we saw Claude Code used a 200k token content. Last week they made it 1 million :)

TEMPERATURE

Temperature controls randomness in token selection.

• Low (0.0–0.3): Predictable, factual, repetitive
• Medium (0.5–0.7): Balanced creativity
• High (0.8–1.0+): Creative, surprising, sometimes incoherent

Most chat interfaces default to ~0.7.

USE CASES

• Drafting: get past the blank page
• Editing: tighten, restructure, change tone
• Brainstorming: generate options, outlines, angles
• Translation: between languages or registers
• Summarization: compress long text to key points

PROMPTING FOR WRITING

Be specific about what you want:

• Audience: “Write this for a general audience”
• Tone: “Casual but informed, not corporate”
• Format: “Three paragraphs, no bullet points”
• Constraints: “Under 200 words”
• Examples: “Match the tone of this paragraph: [...]”

ITERATION IS THE SKILL

The first output is rarely the final output.

• “Make this more concise”
• “The second paragraph is too formal”
• “Keep the structure but rewrite with more specific examples”
• “This is good but cut the last sentence”

Treat it like working with an editor — give feedback, not just instructions.

EXERCISE

Use Claude to write a project description for something you’re working on or want to make.

Think: grant application, residency proposal, or portfolio blurb.

Start by telling Claude about your project in plain language — what it is, why it matters, who it’s for. Then iterate. Push back on the output. Ask for a different tone. Make it shorter. Make it yours.

TIPS

• Don’t accept the first draft — that’s the starting point
• If it sounds like AI wrote it, tell Claude that
• Try giving it an example of writing you like and say “match this tone”
• Specify the audience: a gallery curator reads differently than a tech reviewer
• You can paste in a rough draft and ask Claude to improve it rather than starting from scratch

WHAT ARE SKILLS?

Skills are saved instructions that Claude can follow on command.

• Triggered with a slash command: /my-skill
• Stored as markdown files in your project
• Can include tone, format, rules, examples
• Persist across conversations — teach Claude once, use it forever

GETTING STARTED WITH SKILLS

Anthropic maintains a library of community skills:
github.com/anthropics/skills

The most useful one to start with: /skill-creator

It walks you through building a skill step by step. You describe what you want the skill to do and it generates the markdown file for you. Use it to create your voice & tone skill in the exercise.

WHY THIS MATTERS FOR WRITING

Every time you start a new conversation, Claude forgets your preferences.

A voice & tone skill solves this. Instead of re-explaining “don’t sound corporate, use short sentences, be direct” every time, you encode it once and invoke it whenever you need Claude to write like you.

ANATOMY OF A VOICE SKILL

A good voice & tone skill includes:

• Who you are: your background, what you make
• How you write: sentence length, vocabulary, formality
• What to avoid: corporate jargon, clichés, specific phrases
• Examples: 2–3 samples of writing you’re happy with
• Context rules: when writing for a gallery vs. social media vs. a grant

EXERCISE

Create a Claude skill that captures your writing voice.

1. Gather 2–3 examples of your own writing you like (project descriptions, artist statements, emails, etc.)
2. Ask Claude to analyze your writing style — what patterns does it see?
3. Use that analysis + your own instincts to write the skill
4. Test it: invoke the skill and ask Claude to rewrite something in your voice
5. Iterate until it sounds like you

TIPS

• The “what to avoid” section is often more useful than “what to do”
• Include specific words or phrases you never use
• If Claude’s analysis of your style misses something, correct it
• Try your skill on the project description you just wrote — does it improve?
• You can refine this skill over time as you notice patterns

CHAT vs. AGENT

Chat: You ask, it answers. One turn at a time.

Agent: You give a goal. It plans steps, executes them, reads results, and adjusts. It can use tools — search, read files, browse the web, run code.

RESEARCH TOOLS

• Web search: find recent articles, papers, references
• Deep research: multi-step investigation with synthesis
• Summarization: distill long documents to key points
• Comparison: analyze multiple sources side by side
• Citation: find and verify references

WHEN TO USE AN AGENT

• The task has multiple steps you’d do yourself
• You need information from several sources
• The answer requires reading + synthesizing
• You want a first pass before diving in yourself

Don’t use agents when a single question gets you what you need.

EXERCISE

Take the project description you wrote earlier and get feedback from multiple AI “personas.”

Create agents with different perspectives — each one reviews your description and gives feedback from their point of view. You decide which feedback to act on.

SUGGESTED PERSONAS

• Gallery curator: Is this compelling enough for a show proposal?
• Grant reviewer: Is the impact clear? What’s missing?
• Fellow artist: Does this honestly represent the work?
• General audience reader: Can someone outside the art world understand this?
• Harsh critic: What’s the weakest part of this description?

HOW TO DO THIS

Create agent files in .claude/agents/ in your project:

Each agent is a markdown file with YAML frontmatter (name, description, tools) and a system prompt body that defines the persona.

Example: .claude/agents/gallery-curator.md

AGENT FILE ANATOMY

---
name: gallery-curator
description: Reviews project descriptions from a curatorial perspective. Use when reviewing proposals for exhibitions.
tools: Read, Grep, Glob
---

You are a gallery curator reviewing project proposals. You’ve curated 50+ shows. Be honest and specific. What works? What’s unclear? Would you show this?

INVOKING YOUR AGENTS

• @-mention: Type @ and select the agent from the list
• Natural language: “Use the gallery-curator agent to review my description”
• Chain them: “Use the grant-reviewer then the harsh-critic on my draft”
• Run in parallel: “Use the curator and the general-audience agent in parallel, then summarize”

Compare the feedback — where do they agree? Where do they contradict?

TIPS

• Set tools: Read, Grep, Glob so reviewers can’t edit your files
• The more specific the persona prompt, the better the feedback
• Try making one persona adversarial — it’s the most useful
• Use the feedback to revise, then run the panel again
• Run /agents in Claude Code to create agents interactively

HOMEWORK 1: SCALE YOUR WRITING SYSTEM

You wrote one project description today. Now scale it.

Build a set of tools — skills, agents, or both — that let you write a polished project statement for every project in your portfolio.

Think about what stays the same across projects (your voice, your audience, the format) and what changes (the project itself). How do you encode the consistent parts so you only need to provide the unique details each time?

HOMEWORK 2: ARTIST & PROJECT AUDIT

Use Claude as a research agent to find artists and projects similar to your own work.

• Who is working in the same space as you?
• What projects share your themes, techniques, or medium?
• What are they doing differently?
• What exhibitions, festivals, or publications are showing this kind of work?

Produce a written summary with specific names, links, and your own analysis of where your work fits in this landscape.