AI Advice

Here’s AI advice I generally give people.

How do I use AI better personally?

• Buy a paid AI subscription of ChatGPT, Claude, or Gemini for quality and privacy. The frontier models are much better than the free models, and your data isn’t used to train the models. This is the best $20/month you can spend.
• Have 50 conversations a day with AI. High-usage forces you to find tiny use-cases you’d otherwise ignore.
• If you don’t know what to ask, have it interview you. Ask the AI to interview you to find out what you want, and then do it for you.
• You’ve hired an intern. Don’t treat it like a search engine. It’s as smart as a post-graduate intern - smarter than the average professional in many domains. Give it bigger tasks. Verify its work and correct it (“You missed this part, try again.”)
• Use it for validation. LLMs can make mistakes, but using it to fact-check books, articles, emails, your work, etc. is safe and effective.
• Have AI cross-check AI. Ask it to find all the mistakes it made and give you citations. Have another AI find all the errors. They’re pretty good at that.
• Critique/steelman your work. Ask for counterarguments. Have it roleplay a skeptical customer, boss, or critic; quiz you with hard questions to stress test your work.
• Use emotions as prompts. Unresolved emotions are a group starting point for AI. “I feel anxious…”, “I’m annoyed by …”, “I wish I had…”, etc. are great starting points to
• Ask for easier output. You spend a lot of time validating or implementing AI output. Have AI structure so it’s easy for you to review or do. Your time is the bottleneck. Easy beats best.
• Ask for multiple, diverse, outputs. You don’t know what you want, or what it can do. Ask for 5–10 variations. Ask multiple models. Ask in parallel. Drop the weak ones quickly.
• Use voice mode on mobile to talk to the AI while walking or thinking. “Ramble” at the AI - it can structure your thoughts. This capitalizes on dead time (e.g. commuting) and also lets you dump context and thoughts faster than you type.
• Improve your tools by asking it to optimize your laptop / software, improve settings or configurations, and checking if the results are better.
• Vibe code your own software. As a non-technical person, build apps to solve your own problems. Don’t learn to code. Just tell AI tools what you want and have them build it.
• Have it write code to process numbers. LLMs are were bad at math but good at writing code. Tell it to write code to analyze numbers instead of answering directly.
• Have it rewrite your prompts (meta-prompting). If you aren’t getting the results you want, have it tell you what’s missing and rewrite your prompt for you.

Here are specific ideas you can try:

• Mine your digital exhaust. Don’t delete your “junk” data. Export WhatsApp chats, journal entries, email logs, fitness data, bank statements, etc. and feed them to an LLM. Ask it to find patterns in your behavior, identify your blind spots, or summarize your year.
• Repurpose content and data.
• Read papers, books, and attachments. Have it rewrite in the style of your favorite author (e.g. Malcolm Gladwell) to make dry content more engaging. Add “ELI15” (Explain Like I’m 15) for simplicity.
• Use it when stuck. When you hit a mental wall, use it as a thinking partner. Have it give a first draft, ask it to interview you, ask what an expert or a person you admire would do, or just ramble your thoughts to it.
• Hire an expert. “Hire” it as a personal financial advisor, career coach, relationship counselor, or fitness trainer, and more. For example, hire as a:
  • Doctor. Have it summarize your health history, identify gaps, and suggest questions to ask your doctor.
  • Detective. Ask it to find out what happened to a long-lost friend or what a client has been up to.
  • Financial advisor. Ask it to interview your about your finances, goals, and risk tolerance, then research a personalized investment plan.
  • Relationship Architect. Ask it whom to reach out to, find their interests, what gifts to buy, etc.
  • Teacher. Ask AI to teach, then quiz you. “I want to learn about [Topic]. Explain the basics, then ask me 3 questions to test if I understood it.”

How do YOU use agents?

Here are some of my behaviors in the agent era:

• Prototype the prototype. Sometimes, I’m not even sure what to prototype. I have the agent build something based on very quick, crude early thoughts, then iterate on it. This is as opposed to asking the agent for ideas, filtering them, THEN asking it to build the right prototype. In other words, the reviews are easier when I have a draft rather than an idea. #
• Galleries for ideas. I collect prompts and preview output as galleries. I extend based on usage, but big leaps come when I ask agents to create / extend galleries.
• Audio to analysis. I record calls, transcribe it, and pass it to a coding agent to give the other person what they need – WITHOUT interpreting it. I’ve already put in some hard work into drafting skills, but that may become redundant later. I am mostly getting out of the way of the speed and capability of the agent in solving the problem directly.
• Itch to experiment. When I have a thought, I just have an agent prototype it and run the experiment. With more tools and environments, the space of what it can experiment grows.
• Directional feedback. In areas where I’m not the expert, I tell agents how I feel, how I should feel, how I’ll know if it’s right, and trust the agent’s judgement. #
• Organize context. I record and organize far more data than before (call transcripts, bank statements, phone bills, etc.) to pass to agents. Incidentally, managed digital exhaust is an asset.
• Ask the agent. When I have a question, I ask the agent (not search engine) first. I delegate the research to it, and ask for the answer directly. “Just tell me what to do. Maybe I’ll ask why.”
  • Not just a question. When I have almost any feeling (discomfort, curiosity, etc.) I’m now trained to ask an agent.
  • Many email replies are just copy to Gemini and copy back. It already has the context of my past conversations.

How do I use AI for coding?

Thoughts on technical use of AI (e.g. in when coding with AI)

• Vibe code first. Ask for what you want. Let AI build it. If it works, AND is what you want, AND needs to be maintainable, THEN look at code.
• Non-coders can code. Domain experts (e.g. HR, Finance, etc.) can build their own tools using this way, bypassing traditional IT bottlenecks.
• Use meta-prompting. If you need help, ask AI to write and refine your prompt before you use it for the actual coding task.
• Vibe code end-to-end. Send AI the recording of your client call and ask it to spec, design, build, test, deploy, and monitor. Stay out of the way; review at the end.
• Paste the errors. When code fails, paste the exact error log or a screenshot into the chat. The model is often its own best debugger.
• Code is disposable. Code is an AI compilation artifact. Don’t get attached to it. Scrap and re-tsart.
• The “Two-Strike” Rule. If it fails to fix a bug after two attempts, abandon the thread and restart. It is often faster with a fresh context than to debug a confused model. #TODO May not be true in 2026
• Analyze using code. Ask agents to write code to analyze data. This is more reliable than asking LLMs to analyze directly.
• Use realistic fake data for prototyping. Don’t wait for real data. It’s faster, has no compliance/privacy concerns, and can be as clean/messy as you want.
• Which coding model to use. As of Q1 2026, Claude / Gemini are still good at UI. GPT for rigorous testing.
• Plan unclear tasks. If your idea is vague or might be too complex, ask AI to write an easy-to-review plan. Scan & correct it. THEN implement.
• Maintain reference files. Maintain an up-to-date AGENTS.md (or even README.md) that explains your intent, code, architecture, to the AI. Saves repeated explanations.
• Plan for better models. Models keep improving. In AGENTS.md, skip prompts that work around current model limitations. Focus on what’s true for future models too.
• Generate tests first. For maintainable software, have it define tests first. That makes working code easier. Often, tests can be 2x the code size.
• Use Playwright to verify. Have Playwright take screenshots and inspect DOM elements (e.g. using CDP) to verify frontend work. Saves manual review time.
• Run post-mortems. When it fails, or after any session, ask it to analyze what went well, what didn’t, and how to improve next time. Save these in a SKILL.md.
• Specify developer styles. Ask it to write in the style of a famous developer (e.g. Luke Edwards) or repo (e.g. SciPy) or team (e.g. Astral) that’s apt for the task.

How to drive AI adoption?

Thoughts on the governance & adoption of AI (e.g. organization deployment, challenges, etc.)

• Make using AI easy. Reduce friction. No permission or extra steps required and aligned to current ways of working.
• Show leaders using AI. When teams see leaders using (not talking about) AI, it gives them permission and confidence.
• Security & privacy. Every company has its own white-listed enterprise models (e.g. within Azure, AWS, or Google tenants). Use this. It provides legal cover and data won’t be used for training.
• Keep humans-in-the-loop. Treat AI like an intern that handles 80-90% of the effort, with a human expert for the “last mile” validation.
• Keep updating models. Monitor the ever-shifting “cost-quality frontier” and keep switching to cheaper, better models as they become available. Cost reduces while quality improves.
• Compare accuracy with multiple experts. AI may not match an SME 100%, but one SME may not match another SME either. Check with multiple human experts and see if AI is within the human range of disagreement.
• Use consensus to improve accuracy. Double, triple, or even quintuple-check outputs. If all models agree, accept. Else manual review. This dramatically improves quality while introducing a little human verification overhead.
• Generate code for reliability. Instruct LLMs to write and execute deterministic code (or build models) instead of reasoning in plain text.
• Find AI enthusiasts. Top-down AI mandates build frustration. Prefer “organic adoption”. Find and empower the few enthusiastic “builders” or “power users”.
• Standardize evaluation. You’ll move MUCH faster with evaluation frameworks (like “LLM-as-a-judge”) to score model performance and catch regressions.
• Lay a good data foundation. Convert unstructured documents into (multiple) structured formats. AI output quality depends on input data quality.
• Let anyone build tools. Non-technical “citizen developers” to building their own tools using English, de-bottlenecks IT and dramatically increases prodictivity.
• Prefer less experienced people. Experienced people’s preconceptions hinder AI. Interns and outsiders bring out the best in AI.
• Let the owner drive it. Alice building Bob an AI solution rarely works. Bob building it himself (with Alice’s help) works better.
• Build, don’t plan. When execution is fast and cheap, don’t agonize over the right solution. Build them all. Throw away what doesn’t work.
• Buy, don’t build. Don’t train models. They’re soon obsolete. Build orchestration layers and proprietary data workflows instead.
• Adding is easier than changing. Using AI to improve existing work has a high standard, inertia, and risk. Creating a new workflow or output has less competition.
• Wait for models to improve. Models (and AI products) improve so fast that things that are not possible today will become possible in a few months. Don’t waste time doing what you’ll get for free.
• Wait for the crisis. Real adoption happens when urgency/FOMO temporarily relaxes process. Anticipate that and jump in with demos, clear risk framing, and low-change integration.
• Prototype rapidly. Ask for prototypes in days, not weeks. This builds a culture of “rapid experimentation” and lets you cheaply figure out if it’s worth it.
• Audits make reviews simpler. Ask AI agents to cite sources, provide reasoning, and generate logs. That lets humans to verify how a conclusion was reached.

What skills should I learn?

AI will erode skills - but that s OK for some skills.

• Learn what AI won’t do well even in the future. Practice manually, then use AI for critique & coaching.
• Delegate blindly what AI does well. Use saved time to learn new skills.

Here’s how some industries have dealt with skill erosion:

• Autopilots eroded flying skills - which is dangerous. So we enforce flight simulators. Same for surgical knots (robotic surgery), celestial navigation (navy), manual dosing (nurses).
• Spreadsheets eroded calculation skills. We leveled-up from sums to strategy. Same for CAD, electronic trading, spell-check.
• Photography eroded painting skills. We switched value to impressionism, cubism, etc. Same for vinyl records, luxury watches, craft coffee.
• GPS eroded navigation skills. We accepted this and don’t care much. Same for phone numbers, spelling, mental maths.

Critical skills in the AI era:

• Asking questions. Learn to ask lots of good questions that nudge AI & humans to better results, new horizons. Curiosity helps.
• Thinking of useful things. Learn to quickly discover lots of useful things for yourself and others. AI can execute them fast.
• Validation. AI works fast. Learn shortcuts to compare versions, find mistakes, give feedback - in unfamiliar areas. (Consultants learn this skill well.)
• Accountability. Giving a commitment, standing behind it, knowing how to manage the risk this involves.
• People skills. Prioritize empathy, negotiation, judgment, and communication are less easy to delegate to AI agents.
• Management. Shift from doing the work yourself to managing “teams” of AI interns and agents to handle execution.

Growing skills:

• Storytelling. Learn how to guide AI to deliver compelling narratives that move people.
• Context engineering. Learn what data/context to feed AI and what you can/should skip for the best results.
• Tooling. Learn how to connect things - especially to agentic systems - to give them more execution power.
• Problem breakdown. Learn to breaking problems down into small, logical tasks that people/AI can execute reliably.
• Prototyping. Learn to build (and iterate on) the smallest working solution (using AI agents) ultra-rapidly.
• Data organization. Learn to structure data to make data more analyzable.
• Hard to define skills. Skills that are easy to define are easy to train AI models on. It’s what we can’t even name that’s valuable.

Growing (for a while) skills:

• Learning fast. Learn how to learn faster. You’ll need to learn many subjects quickly (especially to judge AI output). But AI can learn faster.
• Style & art. Learn to guide AI to write, draw, code, etc. in different styles, formats, and approaches for different audiences. But AI can learn these.

Declining skills:

• Coding syntax. AI can write it.
• Factual knowledge. AI can look it up or derive it.
• Domain depth. Unless you are (or can become) a top expert, AI can fill in gaps. Focus on multi-disciplinary knowledge instead.
• Following rules. AI can implement a process better.
• Hard work. Returns are disproportiate. Finding the right problems matters more than solving the lesser problems.
• Building models. AI companies will take care of it.
• Business intelligence. AI can build dashboards, data stories, and more.
• Data wrangling. AI can handle data engineering, modeling, analysis, and visualization.
• Tool expertise. AI can use tools for you.
• Intermediation. AI can translate between groups - e.g. business analysts.
• Originating ideas. AI can brainstorm ideas. Focus on evaluating and selecting ideas based on unique context.
• Drafting from scratch. The ability to write a first draft (code or text) is less valuable than the ability to edit and refine an AI-generated baseline.
• Junior-level execution. Routine “grunt work,” basic summaries, and entry-level analysis are being fully automated by LLMs.

How to develop taste?

See How to develop taste. (But AI can develop taste, too.)

What happens to people when AI takes their jobs?

Here are some paths post-automation. It depends on the industry and individual:

1. Exit: Don’t adapt. There’s no nearby “new task”. You’re unemployed. E.g. bowling pinsetters -> automatic pinsetters; elevator operators; telephone switchboard operators.
2. Downgrade: Serve the machine. Worse job/pay. E.g. textile workers -> power-loom tenders; print compositors -> machine operators; shoemakers -> factory line operatives.
3. Pivot: Focus where automation fails (exceptions, trust, coordination). E.g. bank tellers -> relationship managers; travel agents -> corporate travel desks.
4. Niche: Treat inefficiency as a feature (soul, authenticity). Small market, high margins. E.g. weaving -> artisan textiles; coffee -> baristas.
5. Up-Skill: Master the machine. Become AI-native. Much better job/pay. E.g. human computers -> programmers; draftsmen -> CAD designers; accountants -> advisors.

How can we trust AI when it hallucinates?

How do you trust people who can make mistakes? Treat AI like capable, fallible interns.

• Ask for evidence. Reasons, citations, source links, tests, logs, verifiable checklists, …
• Quintuple-check. Ask multiple AIs. If they all agree, it’s probably right. If they disagree, review manually.
• Ask for code to generate the answer - instead of the answer. Code is more likely right, and easier to verify.
• Make reviews easy. Ask for citations, short & simple summaries, structured output, runnable code, etc.
• Prompt for accuracy. “Never make up an answer.” “If you don’t know, say so.” “Ask me when needed.” “Double-check your work.” “Cite sources.” And so on.

Hallucinations can be a great feature - for creativity, humor, and insight. Don’t always eliminate them. Use as appropriate.

How can I safely share data with AI?

• Pick who you trust. If already trust a provider, e.g. Google, Microsoft, etc., use them. If not, run AI locally or use the techniques below.
• Send schema, run code locally. Send the column names, have AI write code, and analyze it locally. This is safer, and more reliable.
• Anonymize data: Strip or hash PII before sending it to untrusted AI.