Leaning Into AI "Hallucinations" to Make Better Creative Work
Most teams fear that AI will flatten ideas into sameness. Leslie Walsh, head of product strategy and development at Episode Four, is taking the opposite approach. Her team built a tool called RYA and uses large language models to spark weird, unexpected ideas-then anchors them in audience data so clients actually buy them.
The goal isn't faster banner ads. It's faster, stronger platforms that can fuel campaigns across channels.
Data-Grounded Weirdness: How RYA Works
RYA starts with a weekly survey asking Americans how they'd spend extra time or money. It maps preferences across 180 genres and 20 actions like "go on a trip" or "go out to eat."
That turns vague intent into specific passion points: what kind of traveler, what kind of cuisine, what type of experience. Those insights become the spine of idea generation.
From there, the team prompts LLMs-most often Anthropic's Claude for ideation-to generate platform ideas quickly and at volume. The difference: they let the model get a little weird, then shape it with constraints.
Why the Strange Stuff Works
Novel combinations create talk value and fresh mental links. To get them, you adjust prompt strategy and temperature. Higher temperature = more unexpected connections; lower = safer iterations.
Guardrails keep ideas brand-safe and feasible. Strong inputs matter, or you'll get noise. Garbage in, garbage out still applies.
Proof: Ideas Clients Actually Bought
- Financial services: turned financial education into a celebrity-chef cooking series.
- Finance event: transformed product sales materials into custom-pressed EDM vinyls.
- Travel and leisure: launched a cruise ship with a global digital treasure hunt, hiding Google Maps clues across the web.
- Auto lending: created a dating show matching singles with both a car and a partner.
Speed: From 6-8 Weeks to Days
Traditional timeline: brief, debate, ideate, revise-over and over. That eats a month or two.
RYA compresses the loop to days, sometimes a single day. It delivers a big organizing idea you can expand into content, social, product demos, and stunts. Execution comes after the platform is nailed.
Run Your Own "Hallucination-Friendly" Sprint
- Build a lightweight interest graph. Ask your audience how they spend extra time or money. Categorize by genres and actions.
- Pick an LLM for ideation. Claude tends to push more original combinations for early-stage ideas. See Claude.
- Write a prompt scaffold. Include brand constraints, taboos, audience passions, and success criteria.
- Vary temperature. Run batches at 0.3, 0.7, and 1.0. Keep the surprising ones that still fit the brief.
- Set guardrails. Safety topics, compliance notes, feasibility checks, and "must-haves."
- Score ideas fast. Use a 1-5 rubric for novelty, brand fit, earned potential, and buildability. Kill most. Keep a few.
- Workshop live with clients. Show the data backbone, then co-create refinements in-session.
Prompt Patterns That Pull Fresh Ideas
"Combine [audience passion], [brand constraint], and [unexpected format] into a platform that can scale across social, experiential, and partnerships. Output: three platform lines, each with a 1-sentence hook, 3 activations, and a simple KPI."
"Given these audience passions: [list], propose offbeat pairings that still feel useful, delightful, or status-enhancing. Avoid clichés. Keep it legal and feasible within [budget/time]."
Keep It Safe and Sellable
- Brand fit: Map each idea to a core benefit or value. If it's clever but off-mission, cut it.
- Feasibility: Add a line item estimate and a path to prototype in 2-4 weeks.
- Legal/compliance: Bake constraints into the prompt and run a pre-screen checklist.
- Measurement: Define a single primary metric per activation (sign-ups, qualified leads, sampling, time spent).
Where This Is Heading
The most valuable "AI tools" won't be generic. They'll encode your team's instincts: how you brief, how you research, how you make leaps. That's what Episode Four did with RYA-codify domain expertise, then let the model remix it at speed.
AI can be creative. The trick is giving it the right raw material, plus permission to get weird-and the discipline to edit hard.
Sharpen Your Prompt Craft
If you want practical training on prompts, workflows, and creative use cases, explore these resources: Prompt engineering guides.
Your membership also unlocks:
