AI Keynote Slides Need Taste Before Prompts

A friend asked how I make visuals for Both Hands Full and my other talks.

The honest answer is: I made a batch I hated.

That is not me being dramatic. It is the most useful part of the workflow.

I had notes. I had an outline. I had a talk with a point of view. I turned the material into AI image prompts, ran a batch through Rafiki, opened the viewer, and immediately knew the images were wrong. They were polished in the way AI is often polished: confident surfaces, weird little symbols, fake design authority, no pulse.

The batch did not fail because the image model was broken. It failed because the brief was not sharp enough about taste.

It looked like the machine had learned the costume, not the person.

That is the trap with AI keynote slides. The first danger is not that the images look bad. The first danger is that they look finished enough that you start negotiating with them.

"Maybe I can crop this."

"Maybe Canva can save it."

"Maybe this is close enough."

Nope. If the image does not belong in the room, reject it. The model does not get feelings. The human gets taste.

But the really cool part is not just making better slide images.

The cool part is watching an idea move.

My words become a talk spine. The talk spine becomes prompts. The prompts become graphics. The graphics become slides. The slides become a website, a blog post, a guide, a social carousel, a speaker package, a workshop exercise, a festival page, or whatever artifact the work needs next.

That is the thing I am trying to build: not "AI made my slides," but a creative operating system where the original human idea stays traceable as it changes shape.

The First Batch Was Useful Because It Was Wrong

The rejected Both Hands Full batch taught me more than a generic tutorial would have.

It answered the question I had not asked clearly enough: what are these images for?

Were they for the actual deck? For a blog post? For a guide? For a site header? For a private experiment? I had started with "make visuals for the talk," which sounds clear until the outputs arrive and suddenly every image feels like it wandered in from someone else's conference.

That is when the real production question shows up.

An image for a keynote deck has to hold up while I am standing beside it, speaking over it, and letting the room feel something. An image for a KrisKrug.co post has to explain my process and fit the Hope Code lane. An internal guide does not need hero art at all if screenshots, prompt examples, and review notes do the job better.

The bad batch made the boundary visible: generation is not production. Generation is a test. Selection is authorship.

So I marked the batch rejected, kept the provenance, and did not move those images into Canva, the blog, the guide, or the deck. That sounds obvious written down. In the moment, it is the discipline that prevents generic AI slop from sneaking into your work wearing a nice jacket.

A real slide already in the ecosystem. The point of the workflow is not to make every image from scratch; it is to know which artifact belongs to which idea, room, and next use.

Start By Naming The Room

Before I write AI image prompts now, I name the room.

For this post, the room is KrisKrug.co. That means Hope Code, not generic AI futurism and not BC AI institutional graphics. Hope Code is my personal lane: field notes, mycelial maps, analog texture, root lines, marginalia, systems thinking, imperfect human evidence, a little spectral weirdness, and zero patience for corporate tech wallpaper.

BC AI has its own visual language. It is more community-centre, civic, stakeholder-map, public-infrastructure. I like that lane too, but it is a different room. Same root system, different wall paint.

This matters because a style lane is not a pile of adjectives. It is a permission system. It tells the model what kinds of decisions are allowed, what kinds of cliches are banned, and what kind of human I am willing to stand beside.

For the Hope Code pass on this post, I gave the images specific jobs:

• a blog hero showing notes moving into references, prompts, selector review, and Canva finishing
• a reference kit showing that taste comes before prompts
• a selector gate showing rejection as part of authorship
• an editability metaphor showing Canva as the finishing room, not the rescue mission

Those are process visuals. They are not final Both Hands Full deck art. That distinction is small but important. The second batch got closer because I stopped asking for "cool keynote visuals" and started asking for diagrams that served a specific article.

The Talk Spine Still Comes First

For Both Hands Full, the spine is not "AI ethics for creatives."

Too soft. Too generic. Too easy for the model to turn into glowing robots and abstract circuit fog.

The spine is:

You can be pissed about non-consent and still engage.

Left hand: critique.

Right hand: curiosity.

Keep walking.

That sentence does more useful image work than a paragraph of style adjectives.

AI keynote slides fail when they illustrate nouns instead of turns in the argument. "AI ethics" gives you a visual average of the internet. "Left hand critique, right hand curiosity, keep walking" gives you tension, posture, body, and choice.

That does not mean every slide needs a generated image. Sometimes the strongest slide is just the sentence. Sometimes it is a real photo. Sometimes it is a diagram. Sometimes it is a blank beat that lets the room breathe.

The visual workflow has to protect the spine. If the image does not carry the argument, it is decoration. If it looks impressive but nobody can explain why it belongs, it is slop with better lighting.

The Artifact Chain

Here is the pipeline I actually care about.

Raw words turn into a spine.

The spine turns into a script.

The script turns into a slide map.

The slide map turns into prompt jobs.

The prompt jobs turn into graphics.

The graphics turn into slides.

The slides turn into websites, posts, guides, social assets, speaker pages, workshop handouts, and future talks.

That chain matters because it keeps the work from becoming a pile of disconnected AI outputs. Every artifact has a parent. Every prompt has a reason. Every visual can be traced back to the line that made it necessary.

Both Hands Full is a good example. The raw phrase is simple: one hand holds critique, one hand holds curiosity, keep walking. From there, it becomes a talk arc: left hand, right hand, the messy middle. From there, it becomes slide beats: the non-consensual opening, the fears, the transformations, the selector, the responsibility to stay in the room.

Then the same spine travels again.

For the World AI Film Festival in Brazil, that Both Hands Full idea became a keynote for emerging filmmakers: how to keep our souls intact when the machines get really good at making everything. The slide system took on a different style lane: black, white, red, punk zine, photocopy grain, film strips, typewriter urgency, no corporate gloss. The prompt pack did not ask for "cool AI film slides." It asked for specific visual jobs: non-consensual opening, left-hand fears, generation is cheap and taste is not, write for the bot, creative DNA, stay in the room.

One existing slide beat from the broader Both Hands Full / WAIFF lane. This is the kind of artifact that can become a prompt reference, a blog image, a slide example, or a social card, depending on the job.

That is a different kind of prompting. It is not "make me an image." It is "carry this argument into this room in this visual language, and do not betray the source idea."

Once the slides exist, the pipeline does not stop.

The same material can become a KrisKrug.co essay about AI keynote slides. It can become an internal guide for building prompt packs. It can become a workshop exercise where filmmakers write their three-sentence creative DNA. It can become a social post about why generation is cheap and taste is not. It can become a festival recap. It can become a landing page section for a future talk. It can become a better version of the next deck.

That is why provenance matters. I want to know which raw note led to which prompt, which prompt produced which graphic, which graphic made it into which slide, and which slide later became which web artifact.

Without that chain, you just have content.

With that chain, you have a body of work.

The Reference Kit Is The Prompt

Here is the part people skip: before I ask for images, I need a personal style guide for AI images.

Not a brand book in the corporate sense. I mean a working reference kit that says, "This is the neighborhood. This is the weather. This is what I trust. This is what makes me physically recoil."

For a talk, I want 10 to 20 positive references and a handful of negative ones. A past slide that still feels alive. A zine spread with the right density. A map with useful awkwardness. A photo with human scale. A field notebook page. A poster with texture. A diagram that explains without getting precious.

For each reference, I write one line:

• why it works
• what to borrow
• what not to copy

Example:

Positive reference: a messy field notebook spread.

Why it works: dense but human, lots of signals, not over-designed.

Borrow: uneven margins, handwritten arrows, layered evidence, weathered paper.

Do not copy: fake handwriting, fake labels, fake science.

Negative reference: glossy blue humanoid robot on a stage.

Why it fails: corporate futurism, no human stakes, no relationship to my work.

Refusal line: no humanoid robots, no neon circuit boards, no generic innovation theatre.

That little reference kit changes the prompt completely.

Weak prompt:

Make a cool AI keynote slide about turning notes into AI images and editing them in Canva.

Better prompt:

Create a Hope Code process diagram for a KrisKrug.co article about AI keynote slides. Show talk notes becoming a reference kit, then a Rafiki prompt pack, then a selector review gate, then Canva finishing layers. Use analog field-note texture, root-line connections, weathered paper, hand-built map logic, and human editorial restraint. No humanoid robots, no glossy neon, no corporate SaaS dashboard, no fake readable text, no logos.

The second prompt is longer, but length is not the point. The point is that it carries taste, artifact, use, refusal, and visual job. It gives the model less room to decorate and more responsibility to serve the work.

Skip the reference kit and you are not prompting. You are outsourcing taste.

Decide Whether The Slide Needs An Image At All

One of the biggest upgrades is deciding image or no image before generation.

I use a simple mode check.

If the slide is a hard line, make it type. Let the words land.

If the slide is proof, use a real photo, screenshot, artifact, quote, or receipt.

If the slide is a relationship between ideas, use a diagram.

If the slide is a metaphor that cannot be photographed, maybe use generated imagery.

If the slide is an emotional pause, use no image.

That last one matters. Decks do not need to be constantly illustrated. A keynote is not a wallpaper engine. The slide is there to help the room think, not to prove that I know how to use a model.

For Both Hands Full, "left hand critique, right hand curiosity" could become a generated visual. It could also be a clean typographic slide, two real photos, or a hand-drawn diagram. The workflow should consider all of those before spending a generation budget.

The mistake is asking the model to solve a design decision you have not made.

This is also where the artifact chain keeps me honest. A slide that works as a ten-second stage beat might be useless as a blog image. A graphic that explains the workflow in a post might be too busy for a deck. A visual that belongs on a BC AI festival page might feel completely wrong on KrisKrug.co.

Same idea. Different artifact. Different job.

Real photos are part of the chain too. Sometimes the most useful proof is not another generated image; it is evidence that the idea already lived in a room.

Rafiki Runs The Batch. The Viewer Makes The Call.

Rafiki is useful because it gives me a controlled batch workflow: prompt packs, style flags, generations, review, ratings, exports, and provenance.

But Rafiki is not the author. Rafiki is the table where the contact sheets land.

The real work happens in the viewer. I look at each output and ask:

• Does this know what it is for?
• Does it carry the talk spine or just decorate the topic?
• Does it feel like Hope Code, BC AI, or neither?
• Is the composition useful after cropping?
• Is there fake text, fake UI, fake science, or fake symbolism I would have to explain away?
• Would I stand beside this on stage?
• Can Canva finish it without fighting it?

The decision set is simple: approve, approve with Canva, regenerate, or reject.

"Approve with Canva" means the structure is useful but the final artifact still needs human finishing. Maybe the texture works but the labels are nonsense. Maybe the layout is strong but the bottom corner has AI mush. Maybe the metaphor is right but the image needs real text, a better crop, and accessibility checks.

"Reject" means the image is wrong for the work. Not almost right. Not "maybe if we try hard enough." Wrong.

The rejected Both Hands Full batch belongs in that category. The Hope Code process images for this post are candidates. They explain the workflow well enough for a draft, but they are still not sacred. If I look at the final page and decide the post is stronger text-only, that is a valid creative decision.

That is the selector mindset. The generator makes options. The human decides what belongs.

Canva Is The Finishing Room

I do not trust image models with final text.

Not titles. Not citations. Not logos. Not QR codes. Not anything the audience needs to read, scan, trust, or reuse.

If an image survives review, Canva is where I get editability back. I use the Magic Layer style tools when the image has useful pieces. I separate elements. I rebuild text manually. I add real logos manually. I crop for the actual slide, blog post, or social card. I check whether the image still works after the pretty prompt magic has been forced into a real layout.

Canva is not where I rescue off-brand art.

Canva is where approved art becomes usable.

That distinction matters. If the image is wrong, Canva just gives you more ways to waste time. If the image is right but unfinished, Canva turns it into something you can actually ship.

The same is true downstream. A finished slide is not automatically a finished website image. A finished website image is not automatically a good social post. Each artifact needs another human pass: crop, context, headline, alt text, link, caption, call to action, and the humility to say, "This does not belong here."

A Worked Example From Both Hands Full

Here is how I would handle one slide beat now.

Raw note:

You can be angry about non-consent and still stay curious enough to learn the tools. One hand holds critique. One hand holds experimentation. The work is learning to walk with both.

Bad prompt:

Make an AI ethics keynote slide with two hands and futuristic technology.

That prompt is going to drag me toward the mush pile: glowing palms, robot fingers, maybe a floating brain, maybe some fake hologram text. It illustrates the noun "AI" and misses the emotional work.

Better visual brief:

Visual job: show the tension of carrying critique and curiosity at the same time. Stage use: support a spoken line, not replace it. Style lane: Hope Code if this is KrisKrug.co or a personal keynote; BC AI only for a community/institutional version. Composition: two human hands as a field-note diagram, not a sci-fi poster. Left side carries redaction marks, consent receipts, and critique notes. Right side carries a small prototype, seed, or network map. The middle path stays unresolved but walkable. No robot hands, no neon futurism, no readable generated text, no logos, no fake UI.

Selector review:

If it looks like techno-optimism, reject it.

If it looks like anti-AI panic, reject it.

If the hands are grotesque, regenerate or switch modes.

If the metaphor lands but the labels are fake, approve with Canva and rebuild all text manually.

If the sentence is stronger than the image, use the sentence.

That final option is not a failure. It is the workflow working.

A Small Roundup Of The Chain

The related posts and assets are not just "read more" links. They are proof that the idea is already moving through different containers.

Both Hands Full is the source spine: critique in one hand, curiosity in the other, keep walking.

Why Judgment Beats "Creativity" in the AI Era is the taste layer. It explains why the selector matters more than the generator.

Make Culture, Not Content is the downstream rule. The goal is not to spray outputs across channels. The goal is to let the best ideas become useful public artifacts.

You Can't Drink Data is a parallel example of the same posture: critique and curiosity at the same time, with visuals that have to hold tension instead of flattening it.

Punk Rock AI is a reminder that style lanes are contextual. Hope Code is right for this post. Punk zine energy may be right for a film festival room. BC AI may be right for civic/community infrastructure. The workflow has to know the difference.

That is the archive I want: not a pile of disconnected posts, but a traceable body of work where one idea can become a keynote, a deck, a guide, a site section, a public essay, and a new set of references for the next round.

My Practical Loop

The clean version of my keynote visual workflow looks like this:

1. Talk out the messy idea in my own language.
2. Pull out the talk spine.
3. Map the artifact chain: notes, script, prompts, graphics, slides, web, social, guide.
4. Name the immediate room: deck, blog post, guide, site asset, private experiment.
5. Choose the style lane: Hope Code, BC AI, punk zine, or something else.
6. Build a reference kit before writing prompts.
7. Decide which beats need type, photos, diagrams, generated images, or nothing.
8. Write a prompt pack, not one magic prompt.
9. Add refusal lines so the model knows what not to become.
10. Run a small Rafiki batch.
11. Review in the viewer like a selector.
12. Reject anything that does not belong.
13. Bring only approved candidates into Canva.
14. Rebuild real text, logos, layout, and accessibility by hand.
15. Turn the finished slide into the next useful artifact only when it has a real job.

The win is not a folder full of images.

The win is a traceable creative system.

That is why the best AI keynote slides start before the image model: with taste, references, a point of view, and the nerve to reject the shiny thing when it does not feel like the work.

Sometimes the second batch is only good enough for a draft.

That is fine too.

Good enough is how the draft gets honest. Taste is how it gets better.