How LLMs Process Your Marketing Content

Understanding the mechanics of tokenisation, context windows, and training data helps you write better prompts and debug outputs that miss the mark. This lesson is practical, not technical.

In practice: Campaign briefs: 3 hours → 20 minutes

Step 1 of 4

Read

0% done

Why Understanding Mechanics Makes You Better at Prompting

You don't need to understand how a car engine works to drive. But knowing why your car struggles uphill changes your driving habits. The same applies to AI. Understanding two key concepts — the context window and training data — will directly improve your results.

The Context Window: AI's Working Memory

Every AI conversation happens inside a "context window" — the amount of text the model can see at once. Think of it as the model's working memory. If your conversation gets very long, older messages may fall outside this window and the AI effectively "forgets" them.

Practical implications:

For long projects (e.g., rewriting an entire website), break work into sections rather than dumping everything into one conversation
If the AI's responses seem to drift or lose track of earlier instructions, start a fresh conversation with a summary of key constraints
Place the most important instructions (brand voice, audience, format) at the beginning of your prompt, not the end

Training Data and the Cutoff Problem

LLMs are trained on text scraped from the internet up to a specific date. After that cutoff, they have no knowledge of new events, products, or market changes.

What this means for marketers:

Don't ask AI for this year's industry statistics — verify independently
AI won't know about your competitor's product launch from last month
Trend-based content ("what's trending in marketing right now") will be stale

The workaround: Paste current information into your prompt. If you have a recent report, share the key data and ask the AI to incorporate it. This is called "retrieval" and it's highly effective.

Tokens, Not Words

AI processes text in "tokens" — roughly ¾ of a word on average. A 1,000-word document is approximately 1,300 tokens. This matters for:

Understanding pricing (most paid tiers charge per token)
Knowing why very long documents get truncated

The Practical Golden Rules

Front-load your instructions — most important constraints go first
Provide context the AI can't have — paste in current data, your brand guidelines, your audience description
Work in sections for long-form projects — don't try to produce a 3,000-word piece in one prompt

Front-loading instructions — see the difference it makes

The prompt

Write an email about our new feature. We launched a reporting dashboard. It saves time. Our audience is marketing managers. Keep it professional.

Claude's response

Dear [Name],

I hope this message finds you well. I am writing to inform you about our exciting new feature. We have recently launched a reporting dashboard that can help save time. This innovative tool is designed to assist marketing managers like yourself.

Please don't hesitate to reach out if you have any questions.

Best regards,

Quick check

Your AI output suddenly ignores earlier tone instructions after a very long conversation. What is the most likely cause?

Key Takeaways

1The context window is the AI's working memory — long conversations may cause it to "forget" earlier instructions
2Front-load important constraints (tone, audience, format) at the start of every prompt
3Training data has a cutoff date — always verify current statistics and recent events
4Paste current information directly into your prompt to get AI to work with up-to-date data
5Break long projects into sections rather than attempting one enormous generation

Your challenge this week

Apply what you learned in a real task

Take a prompt that produced a weak output recently and rewrite it so the first sentence specifies audience, tone, and format. Run both versions and compare.

Starter prompt · paste into ChatGPT or Claude

Audience: [specific role and company type]. Tone: [adjective, e.g. direct and conversational, not formal]. Format: [structure and length]. Goal: [what the reader should do or think after reading]. [Now write the content: your task here].

Before you practise

Think about the last time an AI output missed the mark. Was the problem the tool — or the order and specificity of your instructions? What would you front-load if you ran the same prompt today?

Was this lesson helpful?

Next step

Put it into practice

You've read the lesson — now apply it in a guided hands-on exercise. It takes about 5 minutes.

Ask the AI Tutor