Lesson 20

Why This Chapter Matters: Moving from Single Prompts to Prompt Systems

By now, you have learned how to craft individual prompts: how to assign roles, structure context, use few-shot examples, and apply chain-of-thought reasoning. But in real professional work, you rarely solve a problem with one prompt. You solve it with a system — a coordinated set of prompts that feed into each other, each one refining, transforming, or validating the output of the previous step.

Consider a real scenario: you are a marketing manager at a mid-sized B2B software company. You need to produce a weekly newsletter, a LinkedIn post, and a short landing page for a new feature. If you write one giant prompt asking for all three, you will get generic, disconnected content. If you write three separate prompts with no shared context, you will get inconsistent tone, repeated facts, and contradictory claims. The professional solution is a prompt system: a sequence of prompts that share a common context block, each with a specific role, format, and validation step.

This capstone chapter teaches you how to design such a system. You will not just learn the theory — you will build one end-to-end for a realistic case, with real tool names, real commands, and a rubric you can use to evaluate your own work.

The Core Architecture of a Prompt System

Every integrated prompt system has five layers. You will see these repeated in every professional deployment, whether you are using ChatGPT, Claude, or an API-based workflow:

• Context Layer: A shared, reusable block of facts about the project — the product, the audience, the brand voice, the constraints. This is written once and pasted at the top of every prompt in the system.
• Role Layer: Each prompt assigns a specific expert role (e.g., "You are a senior copywriter specializing in B2B SaaS").
• Task Layer: The specific deliverable for this step (e.g., "Write a 150-word LinkedIn post").
• Format Layer: The exact output structure (e.g., "Use a hook, three bullet points, and a call-to-action").
• Validation Layer: A final prompt that checks the output against a checklist — tone, facts, length, and brand compliance.

In practice, you will often combine the first four layers into a single prompt, and keep the validation layer as a separate step. This separation is critical: it forces you to review output critically rather than accepting the first draft.

Worked Example: A Content Marketing System for "DataPulse Analytics"

Let us build a real system. The scenario: DataPulse Analytics is a fictional (but realistic) B2B company that sells a dashboard tool for customer support teams. They have just released a new feature called "Sentiment Trends" that automatically categorizes customer tickets by emotion. You need three deliverables:

1. A 200-word product announcement email to existing customers.
2. A 120-word LinkedIn post for the company page.
3. A 300-word landing page section for the feature.

All three must share the same facts, tone, and target audience. Here is how you build the system.

Step 1: Write the Master Context Block

Create a file called context_block.txt in your notes app or a text editor. This block will be pasted at the top of every prompt. It must be specific — vague context produces vague output.

PROJECT: DataPulse Analytics — new feature launch "Sentiment Trends"
PRODUCT: DataPulse is a customer support analytics dashboard. Users connect their helpdesk (Zendesk, Intercom, Freshdesk) and see real-time metrics on ticket volume, resolution time, and CSAT.
NEW FEATURE: Sentiment Trends automatically classifies incoming tickets into Positive, Neutral, or Negative sentiment using a fine-tuned LLM. It shows a 30-day trend line and flags sudden spikes in negative sentiment.
TARGET AUDIENCE: Customer support team leads and operations managers at B2B SaaS companies with 20-200 support agents. They are time-poor, data-literate, and skeptical of AI hype.
BRAND VOICE: Direct, confident, no jargon. Use short sentences. Avoid words like "revolutionary," "game-changing," or "cutting-edge." Prefer "helps you see" over "empowers you to."
KEY FACTS (must not be contradicted):
- Feature is included in the "Professional" plan ($99/agent/month) at no extra cost.
- Available now for Zendesk and Intercom; Freshdesk integration arrives in Q3.
- Setup takes under 5 minutes — no custom training required.
- Accuracy benchmark: 92% agreement with human labeling on a test set of 10,000 tickets.
CALL TO ACTION: For email — "Log in to see your Sentiment Trends." For LinkedIn — link to a demo video. For landing page — "Start your free 14-day trial."

Notice what is included: exact numbers, exact plan names, exact integration names, and a negative constraint (words to avoid). This is what makes the system professional.

Step 2: Create the Email Prompt

Now write the first task prompt. In ChatGPT or Claude, paste the context block, then add the role and task. Here is the full prompt you would use:

[Paste context_block.txt here]

ROLE: You are a senior lifecycle marketing copywriter for B2B SaaS products. You write emails that get opened and clicked by busy operations managers.

TASK: Write a product announcement email to existing DataPulse customers announcing Sentiment Trends.

REQUIREMENTS:
- Subject line: max 45 characters.
- Preview text: max 90 characters.
- Body: 200 words exactly (count them).
- Structure: (1) a one-sentence hook referencing a real pain point, (2) two sentences describing what Sentiment Trends does, (3) one sentence on the 92% accuracy benchmark, (4) one sentence on setup time, (5) a clear call-to-action.
- Tone: direct and confident. Do not use exclamation marks.
- Do not mention the Freshdesk integration (it is not ready yet).

OUTPUT FORMAT:
Subject: [subject]
Preview: [preview]
Body: [full text]

Run this prompt. You will get a draft. Do not edit it yet — you will validate it in Step 5.

Step 3: Create the LinkedIn Prompt

Now create the second prompt. It shares the same context block but has a different role and format. This is how you ensure consistency across channels without copying text.

[Paste context_block.txt here]

ROLE: You are a social media manager for B2B tech companies. You write LinkedIn posts that get comments from support leaders, not just likes.

TASK: Write a LinkedIn post announcing Sentiment Trends.

REQUIREMENTS:
- 120 words maximum.
- First line must be a question that a support team lead would relate to (e.g., "Ever wished you knew how your customers felt before they told you?").
- Include one specific statistic from the context block.
- Include a line about the 5-minute setup.
- End with a soft call-to-action: "Comment 'demo' and we will send you the 2-minute video."
- Do not use hashtags. Do not use emojis.

OUTPUT FORMAT:
Post text only, no preamble.

Step 4: Create the Landing Page Prompt

The third prompt is for the landing page section. This one requires a slightly different structure because it is meant to be read on a screen, not in an inbox.

[Paste context_block.txt here]

ROLE: You are a conversion copywriter specializing in SaaS landing pages. You write for scanners — people who skim headings and bullet points.

TASK: Write a 300-word landing page section for the Sentiment Trends feature page.

REQUIREMENTS:
- Start with an H2 heading (max 8 words).
- Follow with a 2-sentence intro paragraph.
- Then use 4 bullet points, each with a bolded lead-in phrase (e.g., "See trends at a glance:").
- Include the 92% accuracy figure and the 5-minute setup claim.
- End with a 1-sentence call-to-action using the trial link.
- Tone: factual, no hype words.

OUTPUT FORMAT:
H2: [heading]
Intro: [paragraph]
Bullets:
- [lead-in]: [text]
CTA: [sentence]

Step 5: The Validation Prompt

This is the step most beginners skip — and it is the difference between amateur and professional output. After you receive all three drafts, run a single validation prompt that checks them against your context block.

You are a strict quality assurance editor. Below are three pieces of content (Email, LinkedIn, Landing Page) and a context block. Check each piece against the context block and report violations.

CONTEXT BLOCK:
[Paste context_block.txt here]

CONTENT 1 (EMAIL):
[Paste email draft]

CONTENT 2 (LINKEDIN):
[Paste LinkedIn draft]

CONTENT 3 (LANDING PAGE):
[Paste landing page draft]

CHECKLIST:
1. Does any content contradict a fact in the context block? (e.g., wrong plan name, wrong price, wrong integration timeline)
2. Does any content use banned words: "revolutionary," "game-changing," "cutting-edge"?
3. Does any content mention Freshdesk as currently available?
4. Is the email body within 190-210 words?
5. Is the LinkedIn post under 130 words?
6. Does the landing page include the 92% accuracy figure and the 5-minute setup?

OUTPUT FORMAT:
For each of the 6 checks, output: PASS or FAIL, then a one-sentence explanation. If any check FAILS, rewrite that specific sentence or section and show the corrected version.

Run this. If it reports a FAIL, copy the corrected version back into your document. This validation step catches the errors that make AI content feel "off" — subtle contradictions, tone slips, and missing key facts.

Putting It All Together: The Full Workflow

Here is the exact sequence you would follow in a real work session, using ChatGPT (or Claude) in separate chat windows:

1. Open a new chat. Paste the context block, then the email prompt. Run it. Copy the output to a document.
2. Open a second chat. Paste the same context block, then the LinkedIn prompt. Run it. Copy the output.
3. Open a third chat. Paste the context block, then the landing page prompt. Run it. Copy the output.
4. Open a fourth chat. Paste the context block, all three outputs, and the validation prompt. Run it. Fix any FAILs.

Why separate chats? Because each chat has a limited context window. If you run all three tasks in one chat, the model may mix up the roles or carry over formatting from the email into the LinkedIn post. Separate chats keep each task clean.

Common Mistakes

Evaluation Rubric for Your System

When you build your own system, grade it against this rubric. A professional-grade system scores at least 8 out of 10.

• Context completeness (2 points): Does the context block include product name, audience, tone, key facts with numbers, and negative constraints?
• Role specificity (2 points): Does each prompt assign a distinct, relevant expert role?
• Format enforcement (2 points): Does each prompt specify exact output structure (headings, word counts, bullet formats)?
• Cross-consistency (2 points): Do all outputs share the same facts and tone without contradiction?
• Validation quality (2 points): Does the validation prompt check for factual errors, banned words, and format compliance?

Your Practice Task (Under 15 Minutes)

Choose a real scenario you actually work with — a product you sell, a service you offer, or a topic you write about. Write a context block of at least 150 words with: (1) a specific product or topic name, (2) a target audience with a job title, (3) three concrete facts with numbers, (4) two banned words or phrases, and (5) one call-to-action. Then write one prompt (not three) that uses that context block to produce a single deliverable — for example, a 100-word product update for a Slack channel or a 150-word internal memo.

Run the prompt. Then check the output against your context block manually: does it contradict any fact? Does it use a banned word? If yes, rewrite the prompt to add a constraint that prevents it. This exercise takes 10-15 minutes and gives you the foundation for the full three-deliverable system you will build in the next chapter.