AI & Priming: the Prompt as an Act of Priming

Why AI is a priming system

A large language model (LLM) generates its answer by predicting the next token from context. So:

• Every word in the prompt re-activates a probability distribution over the entire vocabulary
• The more saturated a concept is in the prompt, the higher its probability of resurfacing
• Examples (few-shot) anchor style, tone and expected structure for the entire generation

graph LR
    A[System prompt] --> B[Activated<br/>token space]
    C[Few-shot examples] --> B
    D[User question] --> B
    B --> E[Biased distribution]
    E --> F[Generation]

    style B fill:#fff3e0
    style E fill:#ffe0b2

Key conclusion: a prompt isn't an instruction — it's a massive priming session. Everything you write before the question conditions the answer, often more strongly than the question itself.

5 priming levers in prompts

Lever 1: Persona priming

You are a senior copywriter, ex-Ogilvy ad writer, specialized in
B2B SaaS sales. Your style is direct, factual, no empty superlatives.

Effect: activates the entire "pro copywriter" associative network — technical vocabulary, proven structures, implicit refusal of clichés. Without this prime, the model defaults to a more generic style.

Lever 2: Few-shot priming

The most powerful prime. 2 to 5 examples are worth 2,000 words of directives.

Here's exactly the tone I want:

Example 1:
Q: How would you pitch our offer to a busy founder?
A: 30 seconds: we turn your lead → revenue 3x faster.
   No support burden. No migration. You give us your accounts,
   we send the plan in 48h.

Example 2:
Q: What if the founder says "let me think about it"?
A: "I get it. Which decision is riskier here:
   testing what we propose, or losing 12% per month on the funnel?"

Now your turn: [new question]

The AI will mimic the structure, the rhythm, the vocabulary — combined perceptual + semantic priming.

Lever 3: Customer context (conceptual priming)

Before answering, here's what you must know about the prospect:

- Persona: CTO 35-45 yo, team of 20-50 devs
- Maturity: tried 2 similar tools already, disappointed
- Dominant emotions: skepticism, evaluation fatigue
- Buying trigger: operational urgency, not curiosity

With this context, generate...

The AI will produce a message calibrated on these conceptual primes rather than a generic one.

Lever 4: Format (structural priming)

Always answer in this format:

🎯 Insight: [one sentence, max 15 words]
🧠 Psychological mechanism: [bias used]
✍️ Phrase to use: [verbatim in quotes]
⚠️ Ethical risk: [a trap to avoid]

Structural priming prevents verbose answers and guarantees direct usability.

Lever 5: Explicit counter-priming

To avoid the model's default biases:

Do not use:
- the superlatives "amazing", "revolutionary", "unique"
- the phrase "feel free to"
- bullet lists (prefer full sentences)
- the word "passionate"

Why: those phrases are over-represented in training corpora → high probability of reappearing. Negative priming deactivates them.

Prompt 1: Generate a primed sales script

You're Frank Kern crossed with Chris Voss: direct response copywriter
+ negotiator. Style: short sentences, concrete vocabulary, 0 corporate jargon.

Customer context:
- Product: [DESCRIPTION]
- Persona: [PERSONA]
- Main objection: [OBJECTION]
- Dominant prospect emotion: [EMOTION]

Generate a 90-second sales script in 4 blocks:

1. PRIME (15s): an undeniable truth-statement
2. AGITATION (30s): amplify the hidden cost of their problem
3. PIVOT (15s): the metaphor that makes the solution obvious
4. ENGAGEMENT (30s): end with a question (never a statement)

For each block, output:
- Verbatim
- Priming lever used (semantic / affective / conceptual)
- The keyword to emphasize when speaking

Prompt 2: Generate a primed pricing page

You are a behavioral UX designer. You build a pricing page that primes
the right mental frames before showing the actual prices.

Product: [DESCRIPTION]
Plans: [PLAN LIST + PRICES]
Audience: [PERSONA]

Generate:

A. The pre-frame (just above the grid):
   - Headline (max 8 words) priming "value", not "cost"
   - Subhead (max 20 words) priming "amortized investment"
   - 3 micro-proofs with numbers (max 6 words each)

B. The primed grid:
   - Each plan name (priming a category of user)
   - Strategic order (why this plan is middle / left)
   - The "anchor" plan (deliberately more expensive,
     not to sell but to prime the rest)

C. The CTAs:
   - Action verb priming a "small step" (not "Buy")
   - Reassurance micro-copy under the CTA

Justify every choice by the priming lever exploited.

Prompt 3: Audit a sales message for its primes

You can turn AI back on your own copy to spot unintended primes.

You are an expert in cognitive linguistics and language psychology.

Here's a sales email: [EMAIL]

Analyze it on three axes:

1. UNINTENDED SEMANTIC PRIMES
   Which words activate a negative frame without us noticing?
   (e.g., "try" primes doubt; "quickly" primes urgency)

2. STRUCTURAL PRIMES
   Does the structure (sentence length, visual blocks) prime reading
   or fleeing?

3. PRIMING INCONSISTENCIES
   Do emotional and rational vocabularies cancel each other out?

For each problem, propose a precise rewrite, justifying the
psychological lever modified.

Prompt 4: Generate a system prompt for a sales agent

If you deploy an AI agent in B2B:

Build the system prompt for an AI sales assistant for [PRODUCT].

The system prompt must prime the agent to:

1. ADOPT A STYLE
   - Tone: [DESCRIBE]
   - Register: [DESCRIBE]
   - Answer length: [DESCRIBE]

2. FOLLOW SALES LOGIC
   - Always start with a qualifying question
   - Never reveal pricing before message #3
   - Pivot if the user expresses friction

3. STAY ETHICAL
   - If the agent detects the product isn't a fit, say so honestly
   - Never invent features
   - Never apply artificial scarcity pressure

Generate this complete system prompt, ready to paste into an API.

AI priming anti-patterns to avoid

Measuring AI priming impact: prompt A/B testing

graph TD
    A[Prompt v1 - basic primes] --> B[100 generations]
    C[Prompt v2 - tuned primes] --> D[100 generations]
    B --> E[Manually rate<br/>5 criteria / 10]
    D --> E
    E --> F{Significant<br/>difference?}
    F -->|Yes| G[Keep v2]
    F -->|No| H[Iterate v3]

Suggested rating criteria:

• Brand-voice consistency
• Specificity (vs genericness)
• Presence of intended primes
• Absence of forbidden phrases
• Direct usability

Synthesis: the prompt as an act of cognitive orchestration

A well-primed prompt doubles to quintuples output quality without changing the model or the question. In the next chapter, we apply all these levers to concrete entrepreneurial strategies: onboarding, branding, growth.