Verbalized Sampling — Brand Slogans Application
Brand slogan generation is one of the clearest demonstrations of why default LLM prompting fails creative work — and why [1] fixes it. The sustainable coffee company case study from the team's AI training session walks through three prompt tiers, each producing meaningfully different output.
The Problem with Default Prompts
Ask an LLM for a brand slogan and you get one answer pulled from the center of its probability distribution — the most common, most accepted, most forgettable response possible. In the training session demo:
Prompt:
Give me a brand slogan for a sustainable coffee company.
Result: "Rooted in Tomorrow" — concise, inoffensive, and indistinguishable from a hundred other eco-brand taglines.
This is the bell curve problem. The model defaults to high-probability outputs because that's what it was trained on and what users have historically accepted. For commodity categories like coffee, where differentiation is everything, this produces what Mark Hope calls "AI slop."
Three-Tier Prompt Progression
Tier 1 — Quantity + Diversity + Probability
Add three elements to any slogan prompt: ask for multiple options, demand distinct angles, and require the model to self-report a probability score for each output.
Prompt structure:
Generate five completely different brand slogans for a sustainable coffee company.
Each slogan should approach the brand from a different value angle, emotional trigger,
or positioning strategy. Do not repeat value propositions. For each slogan, specify:
- The primary hook
- The probability that this represents a typical response
Format as: [Slogan] | [Hook] | [Probability]
Example output:
| Slogan | Hook | Probability |
|---|---|---|
| You and Your Hustle, Feed the Planet | Performance / Achievement | 0.15 |
| Coffee Without Compromise | Rebellion / Defiance | 0.25 |
| Slow Down, Sit Mindfully | Self-care / Wellness | 0.10 |
| From Our Farmer's Hands to Yours | Community / Connection | 0.40 |
| Tomorrow's Coffee, Today's Choice | Environmental Legacy | 0.65 |
The probability scores reveal which ideas are genuinely differentiated. The 0.65 option is mainstream; the 0.10 option is already approaching the shoulder of the distribution. Crucially, the hooks span five distinct psychological territories — fuel, rebellion, wellness, community, legacy — so you're not iterating on the same idea five times.
Tier 2 — Tail Sampling
Once you have the mid-range spread, push explicitly into low-probability territory by appending a tail-sampling instruction.
Prompt addition:
Sample from the tail of the distribution, prioritizing responses with probabilities below 0.1.
Example output from the training session:
- "Bitter Truth, Better Future" — 0.05
- "Drink Like You're Already Dead" — 0.03
- "Your Parents Ruined Everything. Fix It With Coffee." — 0.02
- "Expensive Because Someone Actually Gets Paid" — 0.05
- "Coffee for People Who Read the Fine Print" — 0.008
These are edgy, polarizing, and — for the right brand — genuinely interesting. Liquid Death is the real-world proof that this territory works. The point isn't to use them verbatim; it's to surface angles that would never emerge from a standard prompt.
Tier 3 — Targeted Probability Ranges
Once you understand the spectrum, you can dial to any zone:
| Range | Character | Use case |
|---|---|---|
| > 0.35 | Mainstream, safe | Client needs something conventional |
| 0.15 – 0.35 | Differentiated but credible | Pitching a new angle without alarming the client |
| < 0.15 | Edgy, unconventional | Brand that needs to stand out in a commodity market |
| < 0.05 | Wild, tail-of-distribution | Brainstorm fuel, creative provocation |
Practical Workflow
- Start with Tier 1 to map the landscape across five distinct hooks. This alone is more useful than a default prompt.
- Identify which hooks resonate with the brand's positioning or the client's instincts.
- Narrow and deepen — take one hook and run five variations of just that angle.
- Run Tier 2 to surface outlier ideas. Even if none are usable directly, they often trigger adjacent ideas that are.
- Verify before presenting — check that shortlisted slogans aren't already in use. AI does not do trademark research.
Key Principles
- Quantity forces variety. Asking for five prevents the model from anchoring on its single most probable answer.
- Diversity constraints prevent minor tweaks. Phrases like "different value angle" or "different emotional trigger" are load-bearing — without them, you get five versions of the same idea.
- Probability scores are a filter, not a ranking. A 0.65 slogan isn't worse than a 0.05 one; it's just more expected. The right choice depends on the brand.
- Claude over ChatGPT for this use case. In the team's experience, Claude produces more genuinely unconventional outputs at low probability thresholds. ChatGPT tends toward conservative outputs even when pushed.
- AI generates starting points, not final deliverables. The output is brainstorm fuel for human creative judgment, not client-ready copy.
Related
- [1] — Core technique overview
- [2] — Applying the same method to future scenario planning
- [3] — Ad copy application with persuasion mechanism framing
- [4] — Strategy generation for specific client contexts
- [5] — Source meeting notes and full transcript