Level 1Lesson 6⏱️ 55 min

Advanced Prompting & Workflows

Move beyond single prompts. Learn to chain prompts, manage context, and build multi-step AI workflows.

TeacherManagerDeveloperAnalystBusinessDoctor

Why Simple Prompts Hit a Ceiling

You've mastered the RACE framework. You can write a good prompt. But what happens when your task is genuinely complex?

Asking an AI to do everything in one prompt is like asking a human to write a novel by just saying "write a book." You need to break it down into steps.

ℹ️

The Ceiling Problem: A single prompt gets you 80% of the way there, but the last 20% is where complexity lives. You can increase prompt detail (longer context), but there's a limit before the AI gets confused or forgets the beginning.

Single Prompt (Won't Work)

Attempting Everything at Once

Write a 5-page research report on climate change impacts on agriculture in sub-Saharan Africa, including 10 peer-reviewed sources, with a literature review, methodology, findings, and recommendations, in academic tone, with proper citations, ready to submit to a journal.

Result: Long, potentially inaccurate citations, mediocre quality because the AI is juggling too many constraints.

Chained Prompts (Will Work)

Step 1: "Find 10 peer-reviewed papers on climate change and agriculture in Sub-Saharan Africa. For each, give me: title, authors, year, and one-sentence summary."

Step 2: "Write a literature review synthesizing these 10 papers. Focus on common themes."

Step 3: "Write the full research report..." (now the AI has context)

The chained approach gives the AI manageable steps. Each step builds on the last. The output quality jumps significantly.

Prompt Chaining: The Power of Steps

Prompt chaining means running multiple prompts in sequence, where the output of one becomes the input to the next. It's how you build AI workflows.

Example: A Research Report Workflow (4 Steps)

Step 1: Research

Find Sources

You are a research librarian. Your job is to find authoritative sources on a topic. I need: 8-10 academic papers or reports on "renewable energy adoption barriers in developing countries." For each source, provide: title, authors, publication year, and a 1-2 sentence summary of the main finding. Format as a numbered list.

Step 2: Synthesize

Analyze the Papers

You are a research analyst. I have these sources: [PASTE OUTPUT FROM STEP 1]. Your job is to identify 3-4 key themes that appear across these papers. For each theme, write 2-3 sentences explaining what the papers found. Do not add your own opinions.

Step 3: Draft

Write the Literature Review

You are an academic writer. Using these themes: [PASTE OUTPUT FROM STEP 2], write a 1000-word literature review on renewable energy adoption barriers. Organize by theme. Use academic tone. Include a brief intro and conclusion.

Step 4: Polish

Refine & Format

You are a copy editor. Edit this literature review: [PASTE OUTPUT FROM STEP 3]. Check for: clarity, academic tone, flow between paragraphs, and proper transitions. Suggest 3-5 improvements and rewrite the full text with those improvements applied.

Notice: Each prompt is focused. Each step is manageable. The AI can do excellent work at each stage because it's not juggling 10 constraints at once.

Context Management: The Token Window

AI models have a "context window" — a maximum amount of text they can process in one conversation. Once you exceed that, the AI forgets the beginning of your conversation.

ℹ️

Current context windows: Claude 3.5 Sonnet ~200k tokens, GPT-4o ~128k tokens, Gemini 2.0 ~1M tokens, Claude Opus ~200k tokens. One token ≈ 4 characters. So Claude can handle ~800,000 characters before hitting the limit.

The Problem: Long Conversations

You're writing a long report. You've been chatting with Claude for 20 messages. The conversation is now 50,000 tokens. You ask a question. Claude forgets details from message 1. It hallucinates. You get bad output.

The Solution: Summarize and Restart

When a conversation gets long (after ~5,000 tokens), do this:

Copy the best output so far.
Start a NEW conversation.
In the first message, paste the output and your context: "Here's what we've done so far. Next step: [new task]"
Continue from there.

💡

When to restart: After big steps in a workflow, paste a summary and continue. This ensures the AI always has clear context and doesn't "forget" your earlier work.

Pro tip: In Claude Projects (we'll cover later), this is built-in. The AI never forgets your context because it's stored persistently. No restarts needed.

Custom Instructions: Persistent Personality

Custom Instructions tell an AI system "here's who I am, here's how I like to work, here are my preferences." Once set, they apply to EVERY conversation with that AI.

ℹ️

Where to find them: Claude → Projects tab (we'll learn this), ChatGPT → Settings → Personalization → Custom Instructions, Gemini → Will be added soon.

Instead of repeating your context in every prompt, you set it once. Forever.

Example: Custom Instructions for a Business Analyst

What a Business Analyst Would Put in Custom Instructions

ABOUT ME: I'm a business analyst at a mid-size fintech company. I work with C-suite, product, and engineering teams. MY PREFERENCES: - Always structure responses with Executive Summary first, then Details - Use data-driven language (avoid "I think" — say "The data shows") - When I ask for analysis, include: Finding, Impact, Recommendation, Next Steps - Use tables for comparisons; use bullet points for lists - Keep explanations technical but accessible (avoid jargon without explanation) MY WORKFLOW: - I often need 1-page summaries and 5-page deep dives for the same topic - I work with large datasets; feel free to ask for clarification on formatting - I present to executives on Friday mornings; flag any uncertain claims MY GUARDRAILS: - Never include proprietary company info in examples - Ask before using email/Slack tone — be formal unless told otherwise

Now, every conversation with that AI automatically starts with that context. No need to repeat yourself. The AI understands your role, your preferences, your workflow, your constraints.

Multi-Turn Strategies: Advanced Conversation Tactics

A "turn" is one message from you and one response from the AI. Smart multi-turn strategies squeeze more value from the AI by structuring how you interact over multiple turns.

Strategy 1: Progressive Drill-Down

Start broad. Get output. Ask for deeper analysis on one piece. Keep drilling down.

Progressive Drill-Down Example

Turn 1: "Summarize the main barriers to renewable energy adoption." Turn 2: "Focus on 'cost.' Why is it still a barrier even as solar gets cheaper?" Turn 3: "You mentioned 'upfront capital.' For developing countries, what policies could reduce this barrier?" Turn 4: "Give me 3 specific examples of countries that reduced upfront barriers with policy."

Each turn builds on the last. You're not re-explaining; you're refining focus.

Strategy 2: Critic Loop

Get a draft. Ask the AI to critique its own work. Refine based on the critique.

Critic Loop Example

Turn 1: [You provide output to edit] Turn 2: "Critique this for: clarity, accuracy, and tone. What's weak?" Turn 3: [AI gives itself critical feedback] Turn 4: "Now rewrite fixing those issues." Turn 5: [AI rewrites with improvements]

This forces deeper thinking and catches the AI's own mistakes.

Strategy 3: Devil's Advocate

Propose an idea. Ask the AI to argue against it. Refine based on weaknesses found.

Devil's Advocate Example

Turn 1: "Here's my proposal for fixing X: [your idea]" Turn 2: "Play devil's advocate. What are the 5 strongest arguments AGAINST this?" Turn 3: [AI gives counterarguments] Turn 4: "How would I respond to each of these?" Turn 5: [AI helps you build counter-counterarguments]

Strengthens your thinking before you present to others.

Strategy 4: Multiple Perspectives

Ask the AI to approach the same problem from different angles in one conversation.

Multiple Perspectives Example

Turn 1: "Analyze this decision from a cost perspective." Turn 2: "Now analyze the same decision from a user/customer perspective." Turn 3: "Now from a risk/security perspective." Turn 4: "Now from an employee/team perspective." Turn 5: "Synthesize all 4 perspectives. What's the strongest recommendation?"

Forces comprehensive thinking. Catches blind spots.

3 Real Workflow Examples by Profession

Example 1: Teacher Creating Differentiated Lesson Materials (4 Steps)

Goal: Create versions of the same lesson for advanced, on-level, and struggling students.

Step 1: Plan the lesson

Step 1

You are an experienced elementary school teacher. I'm teaching fractions to 4th graders. Create a lesson plan that includes: learning objective, 2 concrete manipulatives activities, 1 real-world application, and 1 formative assessment. Keep it to 30 minutes total. Target on-level students (typical 4th grade math).

Step 2: Create advanced version

Step 2

Using this lesson plan: [PASTE STEP 1 OUTPUT]. Modify it for advanced students (working at 5th-6th grade level). Add: one extension activity that introduces multiplication of fractions, and one challenge problem. Keep the same total time.

Step 3: Create struggling student version

Step 3

Using this lesson plan: [PASTE STEP 1 OUTPUT]. Modify it for students who struggle with fractions. Add: a pre-activity reviewing halves/fourths using pizza, more repetition with manipulatives, and a simplified assessment (verbal instead of written). Include specific language to use.

Step 4: Create parent communication

Step 4

Write a 1-paragraph note to parents explaining what we learned about fractions today and how they can practice at home with real objects (pizza, chocolate bars, cookies). Keep it encouraging and specific.

Example 2: Business Analyst Writing Client Report (5 Steps)

Goal: Create a professional report analyzing client data and recommending strategic changes.

Step 1: Analyze the data

Step 1

You are a data analyst. I have client performance data: [PASTE DATA]. Calculate: total revenue, growth rate, top 3 products, top 3 customer segments, churn rate, and NPS trend. Format as a table with findings in plain language.

Step 2: Identify key insights

Step 2

Using this analysis: [PASTE STEP 1]. What are the 3-5 most important patterns? For each, explain the business implication (so what?). Format: Finding → Impact → One-sentence bottom line.

Step 3: Generate recommendations

Step 3

Based on these insights: [PASTE STEP 2]. Propose 3 strategic recommendations to improve revenue and customer retention. For each: What to do, Why (based on data), Expected impact, and Timeline. Be specific and actionable.

Step 4: Create executive summary

Step 4

Write a 1-page executive summary for the client's C-suite. Include: situation (data summary), key insights (3-5 bullets), recommendations (3 bullets), and expected impact. Use confident, business language. No jargon.

Step 5: Build a presentation outline

Step 5

Create a presentation outline (8 slides) for presenting this to the client. Include: Cover slide, Context, Data findings (2 slides), Insights (2 slides), Recommendations, and Next steps. Add bullet points for what to say on each slide.

Example 3: Developer Writing Architecture Documentation (4 Steps)

Goal: Document a new system so other developers can understand and maintain it.

Step 1: High-level overview

Step 1

You are a technical writer. Here's a system architecture: [DESCRIBE SYSTEM]. Write a 200-word overview that a junior developer could understand. Include: what the system does, main components, why it was built this way, and key tech stack choices. Avoid jargon where possible.

Step 2: Component documentation

Step 2

Using this overview: [PASTE STEP 1]. For each component, write: name, purpose (2-3 sentences), responsibilities (bulleted), and dependencies (what it relies on). Format for easy scanning.

Step 3: Data flow diagram description

Step 3

Describe the data flow through this system: [DESCRIBE FLOW]. Explain: how data enters, where it's processed, how it's stored, and how it exits. Use numbered steps. Include potential failure points.

Step 4: Troubleshooting guide

Step 4

Write a troubleshooting guide for common issues in this system. For each issue: name, symptoms, likely cause, and fix. Include where to look in the codebase for investigation.

Hands-On: Build Your First Multi-Step Workflow

🖥️HANDS-ON EXERCISE 1⏱ 5 min

Choose Your Workflow

Pick a real task you do regularly that has 3-5 clear steps.
Examples: writing a proposal, creating a lesson, analyzing data, planning a project, writing a blog post, giving feedback.
Write down the steps in order.

Quick prompt to get started: "Break down [your task] into 4-5 sequential steps, where each step's output feeds into the next."

🖥️HANDS-ON EXERCISE 2⏱ 20 min

Write a Prompt for Each Step

Using the RACE framework, write a prompt for Step 1. Test it in Claude or ChatGPT.
Copy the output from Step 1.
Write a prompt for Step 2. Paste Step 1 output into it. Test it.
Repeat for remaining steps.

💡

Keep each prompt focused. Don't ask the AI to do Steps 1-5 in one prompt. One step per prompt is the rule.

🖥️HANDS-ON EXERCISE 3⏱ 10 min

Document Your Workflow

Create a simple text file or Google Doc called '[Your Task] Workflow'.
List the task name, the overall goal, and the 3-5 steps in order.
For each step, paste your prompt and note how long it took to get good output.
Save this somewhere you can find it later.

This becomes your template. Next time you do this task, you can reuse these prompts.

🖥️HANDS-ON EXERCISE 4⏱ 15 min

Test It Again

Use your workflow on a different instance of the same task (different lesson, different proposal, different data).
Track what changes: Do all prompts still work? Are any tweaks needed?
Update your prompts based on what you learned.

Your prompts get better each time. After 3 iterations, they're really solid.

Multi-Step Workflow Quick Reference

The RACE Framework (Reminder)

R = Role | A = Action | C = Context | E = Expectations

When to Use Chaining

Your task has 3+ clear steps. The output of one step feeds into the next. The task is complex enough that a single prompt confuses the AI.

Prompt Chaining Rules

1. One step per prompt. 2. Paste prior output at the start. 3. Keep each prompt focused. 4. Test each step independently first.

Context Window Management

Restart conversations after ~5k-10k tokens. Summarize key output and paste at the start of the new conversation.

Multi-Turn Strategies

Progressive Drill-Down (go deeper), Critic Loop (self-critique), Devil's Advocate (stress-test), Multiple Perspectives (360 view).

Custom Instructions

Set once, apply forever. Include: role, preferences, workflow patterns, guardrails.

When to Restart

After major steps. When a conversation exceeds 5k-10k tokens. When you need a fresh perspective.

← Level 0 CapstoneCustom AI Tools →