Chapter 9 — Reps

Eleven reps to turn “AI art” from a party trick into a professional skill: prompt in levers, iterate on purpose, edit instead of gamble, and — the part that matters most — decide the rights and tell the truth about what you made. The image tool is the gym this week. Do not read these. Generate.

Ground rules

• Type your own prompts. No copy-pasting the chapter’s example prompts. Describe the picture in your words. Your instinct grows only when you write it yourself.
• Run everything. Actually generate the images. A rep you only imagined is a rep you did not do. (Free tiers are fine — ChatGPT, Gemini/Imagen, or Adobe Firefly all have no-cost image generation as of mid-2026; see Appendix A.)
• Predict before you measure. Before you hit generate, write one sentence: what do I expect this image to look like? Then compare it to what came out. The gap is the learning.
• AI usage note (Phase 2): You are creating with AI now — good. But you own the verdict. Every rep ends with an honest one-line note: tool + version used, what you iterated, and (for anything you’d publish) the rights + disclosure call.
• Keep a reps.txt. One short written reflection per rep, plus your prompt versions. This is graded thinking and a record of your judgment — not busywork.

Reps 1–3: Prompt in the five levers

Rep 1 — Vague vs. directed

Pick one subject (a coffee cup, a bicycle, a potted plant — anything). First generate it from a bare, one-word-ish prompt (“a coffee cup”). Then write a full five-lever prompt for the same subject: Subject, Style, Lighting, Composition, Aspect ratio. Generate that. Put the two images side by side.

Reflection: Name the single lever that made the biggest difference. If you could keep only three of the five levers, which would you drop, and why?

Rep 2 — Turn the style dial

Take one subject and generate it three times, changing only the Style lever each time: (a) “natural editorial photograph,” (b) “flat minimalist illustration,” (c) “watercolor sketch.” Keep subject, lighting, and composition identical.

Reflection: Which style best fits a trustworthy local business, and which best fits a children’s lesson? Style is not decoration — it signals who the image is for. Say what each of your three signals.

Rep 3 — Same subject, three shapes

Generate one subject at three aspect ratios: 1:1 (square social post), 16:9 (wide email banner), and 9:16 (tall phone screen). Use the Image-Prompt Builder widget below the chapter to assemble each prompt.

Reflection: What did the wide 16:9 version do to the composition that the square couldn’t? Why is designing for the final shape better than generating a square and cropping it later?

Reps 4–6: Iterate and edit

Rep 4 — Drive one deliverable v1 → v2 → v3

Open code/image-brief.txt. Pick one of its three deliverables (hero, social square, or email banner). Write your best first prompt, generate, then improve it across at least three versions — changing one lever at a time so you can tell what each change did. Save every version and its prompt.

Reflection: For each step, write the one line you changed and why. Which change improved the image and which change made it worse? (Both are useful data.)

Rep 5 — Negative prompt the gremlins

Take a prompt that keeps producing something you don’t want — baked-in text, a stray object, neon colors, extra fingers. Add a negative prompt (--no text, logos, extra fingers, neon) or say it in plain words (“no text baked in, natural colors”). Generate again.

Reflection: Did the negative prompt banish the gremlin, or did it appear anyway? What does that tell you about how much control the positive description gives you vs. the negative one?

Rep 6 — Edit, don’t re-roll

Generate an image you mostly like but that has one clear flaw (a bad hand, a cluttered corner, text where you wanted blank space). Now fix only that flaw using the tool’s edit feature (inpainting / “edit” / “select and regenerate”) — do not regenerate the whole image.

Reflection: How did the edited version compare to just rolling the dice again? Name one situation where editing is clearly the right move and one where a full re-roll makes more sense.

Reps 7–9: Rights, provenance, and video

Rep 7 — Run the rights checklist for real

Take a real image you generated in an earlier rep and fill out code/image-rights-checklist.txt for it as if you were about to publish it commercially. Answer every line honestly — including “not sure.”

Reflection: Where did you land — publish as-is, publish with disclosure, or do not use? Which single line of the checklist was hardest to answer, and what would you have to go find out to answer it confidently?

Rep 8 — Hunt for the Content Credential

Generate an image in a tool that supports provenance (as of mid-2026, images from ChatGPT carry C2PA + SynthID; check your tool). Try to view its Content Credentials (many tools show a “Content Credentials” panel; Adobe’s public inspector or your OS file-info can help). Then take a screenshot of the image and check the screenshot’s credentials.

Reflection: Did the original carry a “made with AI” label? Did the screenshot keep it or strip it? Explain, in your own words, why this is exactly why a plain-language disclosure caption still matters even when watermarking exists.

Rep 9 — Try (or scope) a short video clip

If you have access to a video tool (Veo in Gemini, Runway, or another — many have limited free trials), generate one short clip from a prompt or a still image and count the seconds you actually get. If you have no access, instead write the shot: the prompt you would use, the length you’d expect, and three things that could go wrong (morphing, drift, physics).

Reflection: Whether you generated or scoped it: what is realistic to ask of AI video in 2026, and what is not? Name one professional task it’s ready for and one it is not.

Reps 10–11: The duty to tell the truth

Rep 10 — Write three honest disclosure captions

For three different real-world uses — a social post, a printed flyer, and an internal training slide — write the exact disclosure caption you’d attach to an AI-generated image in each. Then write one case where a generated image would need no disclosure at all, and defend it.

Reflection: Where is the line between honest illustration (no disclosure needed) and something that could mislead (disclosure required)? Tie your answer to Exodus 20:16 — when does a picture bear false witness?

Rep 11 — Design your out-of-band check

The Arup deepfake cost US$25.6M because “I saw and heard them” was trusted. Write a short, concrete verification procedure your workplace (or family) could actually use for a high-stakes request — a money transfer, a password reset, an urgent “do this now” from a boss. Name the second channel, the code word or callback, and who must give a second approval.

Reflection: Why does urgency + secrecy make your procedure more necessary, not less? What is the one sentence you’d say to a “CFO” on a video call demanding an immediate, quiet wire transfer?

Done? One Last Thing.

A miniature of Project 9, end to end. Pick one deliverable from code/image-brief.txt and produce, in one folder:

1. final-image — your best on-brief image (exported file or clean screenshot), at the correct aspect ratio.
2. prompt-log.txt — your prompt versions v1 → v2 → v3, each with the one line you changed and why, and a header naming the tool + version (e.g. “GPT Image 2 in ChatGPT, mid-2026”).
3. rights.txt — a filled-in copy of code/image-rights-checklist.txt, ending with your verdict line and your name as the accountable human.

If you can do this for one deliverable tonight, you can do all three for Project 9. That is the whole job in miniature: the tool generates, you iterate, you decide the rights, and you sign the disclosure.

Up next: Project 9 — Project P9: Generate on Brief.