In the video “You’re not using AI like THIS“ (an interview between Reid Hoffman and AI engineer Parth Patil), they discuss several advanced prompting strategies that move beyond simple “question and answer” interactions. The core philosophy they advocate for is treating AI not just as a search engine, but as a high-bandwidth collaborator that requires rich context and interactive dialogue.
I asked Gemini 3 Pro the following:
I'm trying to get a lot better at constructing more effective prompts for using Gemini. This video talks about several techniques I'd like to understand better. Please educate me further on what they discuss. Include example prompts to illustrate your points. Here's the video: https://youtu.be/cg3Lsr8m2ZQ?si=sGUwPA6ixGM3bnpGGemini’s response follows below:
Here are the key techniques discussed, along with educational explanations and example prompts to help you use them with Gemini.
1. The “Interview Me” Protocol (Meta-Prompting)
Timestamp: [05:30]
Most people rush to ask for a solution immediately. Parth Patil argues that we often don’t even know the full scope of our own problems yet. Instead of asking for the answer, you should ask the AI to help you define the problem first.
- The Concept: Force the AI to gather the necessary context from you before it attempts to solve anything. This mimics how a senior consultant or engineer would interact with you—they wouldn’t just start coding; they would ask questions first.
- Why it works: It prevents “hallucinated” or generic advice by grounding the AI’s response in your specific constraints (budget, timeline, personal preferences) that you might have forgotten to mention.
Example Prompt:
“I want to [insert goal, e.g., ‘launch a newsletter’ or ‘refactor this code’]. Describe the problem back to me, but before we begin, interview me. Ask me clarifying questions until you feel you have enough context to provide the best possible solution. Ask the questions one by one.”
2. Extreme Role Assignment (Persona Swarms)
Timestamp: [03:23]
While “Act as an expert” is common advice, Patil suggests taking this much further by using extreme or opposing personas to stress-test your ideas. He mentions creating “100,000 unique expert personalities” to cover every angle.
- The Concept: You can simulate a “board of directors” or a specific user base. You don’t just want an answer; you want a perspective.
- Why it works: An “optimist” will give you different advice than a “skeptical VC.” By asking for specific emotional or professional lenses, you break the AI out of its default “neutral assistant” tone.
Example Prompt (The Skeptic):
“Pretend you are the most skeptical Venture Capitalist imaginable. I am going to pitch you my idea. I want you to find 10 specific, harsh reasons why this will fail. Do not be polite; be critical.”
Example Prompt (The Swarm):
“I have a problem regarding [topic]. Generate 5 distinct expert personas relevant to this (e.g., a data scientist, a behavioral psychologist, a exhausted parent, etc.). Have each of them write one paragraph giving their unique perspective on how to solve it.”
3. “Voice Pilling” (High-Bandwidth Context Dumping)
Timestamp: [10:46]
Patil describes “Voice Pilling” as perhaps the most powerful technique. This involves using voice-to-text to “ramble” at the AI for 5–10 minutes.
- The Concept: Typing filters your thoughts; you tend to edit yourself for grammar and structure, which leads to “low bandwidth” communication. When you speak (or type in a stream-of-consciousness style), you provide a massive amount of messy but valuable context.
- Why it works: AI is incredibly good at structuring unstructured data. It acts as an editor that can sift through your “brain dump” to find the gold.
- How to do it in text: If you aren’t using voice, you can simulate this by typing without stopping, ignoring typos, and just letting your thoughts flow.
Example Prompt:
“I am going to provide a stream-of-consciousness ‘brain dump’ about a project I am stuck on. It will be messy, unorganized, and full of typos. Please listen to everything, organize my thoughts into a coherent structure, and then tell me what key piece of information I am missing.”
4. The Multi-Agent Orchestration (Manual Swarm)
Timestamp: [29:26]
Patil demonstrates using different AI models (Claude, Gemini, ChatGPT/Codex) simultaneously for different parts of a project.
- The Concept: Treat different AIs like specialized employees.
- Gemini: Use it for large context tasks (e.g., “Read my entire blog history and suggest new topics” [29:58]).
- Claude: Use it for nuance and coding architecture.
- ChatGPT/Codex: Use it for quick execution or web browsing.
- Educational Application: You don’t need fancy software to do this. You can simply open three tabs and assign them different “jobs” for the same project.
Example Prompt (for Gemini, leveraging its large context window):
“I am pasting the transcripts of the last 10 meetings I’ve had. Read all of them and create a psychological profile of my management style, then suggest 3 specific improvements.”
Summary of the “Mental Shift”
The video emphasizes that the biggest barrier to effective prompting isn’t technical skill, but ego [07:07]. Users often try to “steer” the AI too much because they think they already know the answer. The most effective prompters are “humble” enough to say, “I don’t know what the right questions are—you tell me.”