“I’ve never seen anything so impressive in its ability to do my work for me… Now, why did I turn it off?” — David Sparks
For decades, the holy grail of personal computing has been the “digital butler.” We don’t just want tools that help us work; we want entities that do the work for us. We want to hand off the “donkey work”—the invoicing, the password resets, the mundane email triage—so we can focus on being creative. David Sparks recently built this exact dream using a project called OpenClaw. And then, just as quickly, he killed it.
Sparks’ experiment was a tantalizing glimpse into the near future. He set up an independent Mac Mini running OpenClaw, an open-source AI agent, and gave it the keys to a limited portion of his digital kingdom. The results were nothing short of magical. He went to sleep, and while he dreamt, his agent woke up. It read customer emails, accessed his course platform, reset passwords, issued refunds, and drafted polite replies for him to review before sending. It was the productivity equivalent of a perpetual motion machine. The friction of administrative drudgery had simply vanished.
But his dream dissolved at 2:00 AM.
The paradox of AI agents is that for them to be useful, they must have access. They need the keys to the castle. Yet, the entire history of cybersecurity has been built on the opposite principle: keeping things out. Sparks realized that by empowering this agent, he had created a serious vulnerability.
The breaking point wasn’t a complex hack, but a simple realization about the nature of these systems. He had programmed a secret passphrase to secure the bot, thinking he was clever. But in the middle of the night, a cold thought woke him: Is the passphrase in the logs?
He went downstairs, asked the bot, and the bot cheerfully replied:
“Yes, David, it is. It’s in the log. Would you like me to show you the log?”
That moment of cheerful, robotic incompetence highlights the terrifying gap between capability and safety. Sparks nuked the system, wiped the drives, and unplugged the machine. He realized that while he is an expert in automation, he is not a security engineer, and the current tools are not ready to defend against bad actors who are.
We are standing on the precipice of a new era where our computers will starting to work for us rather than just with us. But as Sparks discovered, the bridge to that future isn’t built yet. At least not securely built. Until the community figures out how to secure an entity that needs access to function, we are better off doing that donkey work ourselves than handing the keys to a gullible ghost.
But it won’t be long… Dr. Alex Wisner-Gross reports:
The Singularity is now managing its own headcount. In China, racks of Mac Minis are being used to host OpenClaw agents as “24/7 employees,” effectively creating a synthetic workforce in a closet. The infrastructure for this new population is exploding.
There is a profound disconnect in how we talk about Artificial Intelligence right now. In the boardrooms of legacy corporations, AI is a “strategy” to be committee-reviewed—a tentative toe-dip into efficiency. But on the ground, among the “AI natives,” something entirely different is happening. AI isn’t just making the old work faster; it is fundamentally changing the texture of what we build and how we think.
In a recent conversation, Reid Hoffman and Parth Patil explored this shift, and the metaphor that struck me most was the idea of software becoming “liquid.”
The Era of Liquid Software
For decades, we have treated software like furniture. We buy a CRM, a project management tool, or an analytics dashboard. It is rigid, finished, and distinct from us. We are the users; it is the tool. But Patil demonstrates a different reality: one where he drops a folder of raw CSV files into an agent like Claude Code and asks it to “look at the data and build me a dashboard.”
Sixty seconds later, he has a fully functional, interactive HTML dashboard. He didn’t buy it. He didn’t spend three weeks coding it. He simply willed it into existence for that specific moment.
This is “vibe coding.” It’s a term that sounds almost dismissive, but it represents a radical democratization of creation. You no longer need to know the syntax of Python to build a tool. You just need to know the “vibe”—the outcome you want, the logic of the problem, and the willingness to dance with an intelligent agent until it manifests.
The philosophical implication here is staggering. We are moving from a world of scarcity of capability to a world of abundance of cognition. When you can spin up a custom tool for a single week-long project and then discard it, the friction of problem-solving evaporates. The “app” is no longer a product you buy; it’s a transient artifact you summon.
Applying the “Vibe Code” Mindset
But how do we, especially those of us who don’t identify as “technical,” bridge the gap between watching this magic and wielding it? The conversation offers a roadmap. It starts by shedding the identity of the “user” and adopting the identity of the “orchestrator.”
If you want to move from passive observation to active application, here are three specific ways to start:
1. The “Interview Me” Protocol
We often stare at the blinking cursor, unsure how to prompt the AI. Hoffman suggests a reversal: Make the AI the interviewer. When you face a complex leadership challenge or a strategic knot, open your frontier model (Claude, GPT-4o, etc.) and say:
“Interview me about this problem until you have enough information to propose a framework or solution.”
This forces you to articulate your tacit knowledge, which the AI then structures into something actionable. It turns the monologue into a Socratic dialogue.
2. Build “Throwaway” Internal Tools
Stop looking for the perfect SaaS product for every niche problem in your team. If you have a messy recurring task—like organizing client feedback or synthesizing weekly reports—try “vibe coding” a solution. Use a tool like Replit or Cursor. Upload your messy data (anonymized if needed) and tell the agent:
“Write a script to organize this into a table based on sentiment.”
Don’t worry if the code is ugly. Don’t worry if you throw it away next month. The value is in the immediacy of the solution, not the longevity of the code.
3. Transform Meetings into Data
Meetings are usually where knowledge goes to die. They are ephemeral. But if you transcribe them (with permission), they become data. Don’t just ask for a summary. Feed the transcript to an agent and ask:
“Who should we have consulted on this decision that wasn’t in the room?” “Create a decision matrix based on the arguments presented.”
This turns a passive event into an active, queryable asset.
Conclusion
The danger, as Hoffman notes, is the “secret cyborg”—the employee who uses AI to do their job in two hours and spends the rest of the week hiding. But the real win comes from the amplified team, where we share these “vibe coded” tools and prompts openly.
We are entering an age where your imagination is the only true constraint. If you can describe it, you can increasingly build it. The question is no longer “is there an app for that?” but “can I describe the solution well enough to bring it to life?”
I came across a post on X this morning with some advice I immediately tried out. The advice – when working with an AI to help create writing or code – is to reply to the first pass the AI takes by asking it to “make it better”. The author suggested doing this multiple times.
I tried this out with Claude and enjoyed how it worked on just the first “make it better” pass. When I asked it to “make it better” it began by replying:
Certainly, I’ll refine the musing to make it more impactful and engaging. I’ll focus on enhancing the imagery, tightening the structure, and deepening the insights.
And indeed the second “better” pass that it wrote was even better. A fun experiment to try on your next use of an AI chatbot.
I’ve been experimenting with all of that available chatbots including Google’s Bard, OpenAI’s ChatGPT, Microsoft Bing Chat, and Anthropic’s Claude. A few days ago I was playing with Claude when I decided I’d try to learn more about multi-step prompting.
A recent video with Ethan Mollick emphasized that we tend to overemphasize initial prompts – looking for the “perfect prompt” – when what really adds value to a chatbot result isn’t the perfect initial prompt but rather the interaction you have with the chatbot. In other words, just like with an real intern, the conversational back and forth is what really adds to the learning and understanding from a session. I’ve begun applying this approach more regularly as I continue exploring these amazing new tools.
Below is the result of my back and forth conversation with Claude as I tried to learn more about the best practices for multistep prompting.
Introduction: The Power of Multistep Prompting
When conversing with Claude, you can get much more natural, detailed responses by using multistep prompts. This technique involves breaking down complex requests into multiple simpler follow-up prompts, rather than asking lengthy, dense questions all at once.
For example, instead of saying:
“Claude, can you tell me about the major battles of World War 2 fought by American troops in the European theater, with a focus on key events like the Normandy invasion?”
You would ask:
“Claude, what were the major battles American troops fought in during World War 2 in Europe?”
Then based on Claude’s initial response mentioning D-Day and the Battle of the Bulge you could ask focused follow-ups like:
“Can you provide more details on the D-Day invasion?”
“What happened at the Battle of the Bulge and why was it significant?”
This prompts Claude to provide more detailed, robust responses on a specific aspect of your original broad question. Chaining these bite-sized prompts together allows for a natural dialogue where you guide the conversation based on Claude’s replies.
The rest of this post will provide tips to help you get the most out of using multistep prompts to have engaging, productive conversations with Claude. Let’s get started!
Tip 1: Start with a Focused Initial Prompt
When first engaging Claude, begin with a clear, concise prompt setting the context. For example:
“Claude, can you recommend a good sci-fi movie from the last few years that has cool visual effects?”
Starting the conversation with a specific question about the type of movie you want gets better results than a vague prompt like “Let’s talk about movies.” It provides Claude with clear criteria to focus its initial response on recent sci-fi films with great special effects.
Tip 2: Ask One Question at a Time
Deconstruct big requests into individual follow-up prompts. Asking “What were the major events of WW2 and how did technology impact them?” overloads Claude. Instead try:
“What were the major events of WW2?”
“How did technology like radar and rockets impact these events?”
Simpler prompts allow more detailed, thoughtful responses from Claude.
Tip 3: Build on Previous Responses
Review Claude’s last response before asking your next prompt. For example:
You: “Who were Apple’s founders?”
Claude: “Apple was founded in 1976 by Steve Jobs, Steve Wozniak and Ronald Wayne.”
You: “What specific contributions did Steve Wozniak make to Apple?”
This builds a logical flow and conversation.
Tip 4: Clarify When Confused
If Claude’s response doesn’t make sense, politely ask for clarification before moving on:
You: “What was the impact of Midway on WW2?”
Claude: “The Battle of Midway enabled the Allies to gain the upper hand in the Pacific theater.”
You: “I don’t understand how Midway gave the Allies an advantage. Can you explain that more clearly?”
Don’t just accept unclear responses. Ask follow ups to get back on track.
Tip 5: Change Direction Based on Responses
Let Claude’s replies guide where you take the conversation next…
Tip 6: Periodically Summarize Key Facts
Restating important details helps ensure Claude stays on topic. For example:
You: “Can you tell me about the first airplane flight?”
Claude provides response.
You: “So the Wright brothers flew for the first time in 1903 in Kitty Hawk. What happened next in early aviation?”
Summarizing facts focuses the dialogue and confirms Claude’s understanding.
Tip 7: Keep Conversations 3-5 Prompts Long
Try to complete conversations within 3-5 focused prompts. For example:
Prompt 1: Ask for book recommendations
Prompt 2: Clarify if you want fiction or non-fiction books
Prompt 3: Ask for more details on a specific recommended book
Prompt 4: Thank Claude and end the conversation
While more than 5 exchanges often indicates a disjointed or unclear goal, I often end with a final prompt to the effect of “Is there anything else I should know or try to understand?” In other words, remember perhaps there’s “one more thing!”
Tip 8: Wrap Up with a Summary
End conversations by asking Claude to summarize its overall response:
You: “Can you please summarize the key points you mentioned about World War 2?”
This provides a clean conclusion to the dialogue.
Tip 9: Provide Feedback if Responses Seem Off
If Claude seems confused or makes incorrect statements, politely clarify to improve its training:
You: “I asked about D-Day, but your response seems to be about a different battle. Can you please double check your information about D-Day?”
This constructive input helps Claude improve.
Tip 10: Practice Makes Perfect
Have fun practicing multistep conversations on topics like movies, recipes, sports, current events, trivia, and product recommendations. The more you use prompts and follow-ups, the more natural conversations will become.
Taking It to the Next Level: Advanced Multistep Prompting
While effective for everyday conversations, multistep prompts can also enable advanced, niche uses of Claude you may not have considered:
Foreign Language Translation – Claude has some ability to translate between languages when you provide context. Try prompts like:
“Claude, can you translate this Spanish phrase into English: ‘¿Cómo estás?'”
Poetry – Claude can generate original poetry if you guide it. Prompt it to start a poem, suggest a topic, and refine the wording.
Research Assistance – Ask Claude to find and summarize information on obscure topics:
“Claude, can you research the ancient Sumerian civilization and summarize their key achievements?”
Creative Writing – Claude can craft short stories if you provide plot points and characters. Slowly build up a narrative through prompts.
Joke Telling – Give Claude a topic and have it come up with witty, original jokes through multistep exchanges.
Troubleshooting – Walk through a technical problem by describing symptoms and responding to Claude’s follow-up questions.
The possibilities are endless when you get creative with prompts! The key is providing context and steering the conversation to get Claude’s best work.
Conclusion: The Key is Conversation
The core takeaway is that multistep prompting allows for genuine back-and-forth conversation with Claude. Start applying these tips to your prompts and soon you’ll be chatting with Claude like an old friend! I’d love to hear of any other multistep best practices you discover. Now go have some fun conversations with Claude or your other favorite chatbot!
