3.1.2: Understanding the Basics

• Time to Complete: 15 minutes
• Prerequisites: API key set up (Module 3.1.1)

Start this module in Claude Code: Run /start-3-1-2 to begin the interactive experience.

Overview

Module 3.1.2 teaches you the mechanics of image generation - how the system works and what you can control. You’ll understand the generate() function, learn about aspect ratios and resolution, and master the art of iteration.

Key takeaway: You don’t need to memorize parameters or write code. Describe what you want naturally, and Claude picks smart defaults. But understanding what’s possible helps you get better results.

How Generation Works

When you ask Claude to generate an image:

1. You describe what you want in natural language
2. Claude translates your request into API parameters
3. The generate() function sends the request to Gemini
4. Gemini generates the image (10-15 seconds)
5. The image saves to your outputs/ folder
6. Claude tells you where to find it

You never touch the API directly. Claude handles everything.

The generate() Function

All image generation flows through image_gen.py. Here are the key parameters:

Two Ways to Work

Option 1: Let Claude decide

Generate a professional headshot of a product manager

Claude picks sensible defaults (1:1 for headshots, 1K for drafts).

Option 2: Specify what you want

Generate a professional headshot, 16:9 aspect ratio, 2K resolution

Claude honors your explicit requests.

Both approaches work great. Start with Option 1, get specific when needed.

Aspect Ratios

Aspect ratio is the shape of your image. Choose based on where you’ll use it.

Quick Reference

• Presentation slide? → 16:9
• Social media post? → 1:1 or 4:5
• Phone mockup? → 9:16
• Website hero? → 16:9 or 21:9
• Profile picture? → 1:1

Resolution

Resolution determines size and detail level. It does NOT affect creative quality - just pixel dimensions.

Resolution Strategy

Use 1K while iterating. It’s faster, costs the same, and lets you explore more quickly.

Use 2K for final versions. Once you’re happy with the creative direction, regenerate at higher resolution.

Use 4K only for print. Unless you’re printing at large scale, 4K is overkill.

Iteration: The Core Workflow

Iteration is the most important concept in image generation. Instead of hoping to get it right on the first try, you refine step by step.

Why Iteration Works

Gemini is a “thinking model” - it maintains context across the conversation. When you say “make it bluer,” it knows what “it” refers to and what you’ve discussed before.

Single-shot approach (frustrating):

Generate the perfect image → Hope it's right → Start over if not

Iterative approach (effective):

Generate first draft → "Add more contrast" → "Move the text higher" → "Perfect"

How to Iterate

After Claude generates an image, just ask for changes:

• “Make the background darker”
• “Add a subtle shadow”
• “Change the text to say ‘Launch Day’”
• “Make it feel more professional”
• “Try a warmer color palette”

Claude continues the session with Gemini, and your changes build on the previous image.

When to Start Fresh

Sometimes iteration isn’t the right move:

• Major direction change → Start fresh with new_session()
• Completely different subject → Start fresh
• Want to explore alternatives → Generate variants (covered in 3.1.3)

Tell Claude “start a new session” or “let’s try something completely different” and it will begin fresh.

Sessions Explained

A session is a conversation with Gemini that maintains context. Here’s how it works:

Within a session:

• Gemini remembers previous generations
• You can reference “the image” or “it”
• Edits build on previous versions
• “Thought signatures” preserve reasoning

Between sessions:

• Fresh start
• No memory of previous work
• Good for new projects or directions

Session Management

Claude handles sessions automatically, but you can control them:

Pro tip: Sessions work best for linear refinement. If you want to explore multiple directions, use variants (covered in 3.1.3).

Practical Examples

Example 1: Presentation Graphic

You: “Create a hero image for a presentation about AI productivity”

Claude generates a 1:1 image at 1K resolution.

You: “Make it 16:9 for my slides”

Claude regenerates with correct aspect ratio.

You: “Add text that says ‘AI for PMs’”

Claude adds the text overlay.

You: “This is perfect, regenerate at 2K”

Claude produces the final high-resolution version.

Example 2: Quick Exploration

You: “Generate a user persona portrait - corporate vibe”

Claude generates first version.

You: “Try a more casual look”

Claude refines the style.

You: “Actually let’s start fresh - try illustrated style instead”

Claude starts new session and generates illustrated version.

Best Practices

Do:

• Start at 1K resolution for faster iteration
• Be specific about changes - “make the sky more orange” beats “improve it”
• Let Claude pick defaults when you don’t have strong preferences
• Build incrementally - small changes are more predictable

Don’t:

• Don’t start at 4K - you’ll waste time on images you’re going to change
• Don’t make multiple changes at once - iterate one thing at a time
• Don’t be vague - “make it better” gives Claude nothing to work with
• Don’t abandon good images - iterate instead of starting over

Troubleshooting

Changes aren’t being applied

• Make sure you’re being specific about what to change
• Try rephrasing: “change the background color to navy blue” instead of “different background”
• The session may have gotten confused - start fresh

Aspect ratio looks wrong

• Verify you asked for the right ratio (16:9 vs 9:16 is a common mix-up)
• Some compositions work better in certain ratios - Claude may suggest alternatives

Image quality seems low

• Check resolution - you may be at 1K (which is fine for drafts)
• For final outputs, explicitly ask for 2K resolution

Generation is slow

• 4K takes ~45 seconds, which feels slow but is normal
• Poor internet can add latency
• High API load can cause delays

Quick Reference Card

Aspect Ratios:
  1:1   → Square (profiles, icons)
  16:9  → Landscape (presentations)
  9:16  → Portrait (stories, mobile)
  4:5   → Tall (Instagram feed)

Resolution:
  1K → Fast drafts
  2K → Final outputs
  4K → Print only

Workflow:
  1. Generate at 1K
  2. Iterate until happy
  3. Regenerate at 2K for final

What’s Next?

You understand the mechanics. Now it’s time to learn the art.

Module 3.1.3 teaches the Golden Rules of prompting - how to write descriptions that get amazing results. You’ll also learn about reference images and generating variants.

Interactive track: Type /start-3-1-3

Resources

• Gemini Image Generation Documentation - Parameters, aspect ratios, resolution options
• Gemini 3 Developer Guide - Thought signatures, multi-turn sessions

About This Course

Created by Carl Vellotti. Check out The Full Stack PM for more PM builder content.

Source Repository: github.com/carlvellotti/claude-code-pm-course