After working with AI long enough, I’ve realized the core philosophy boils down to two words: convergence and divergence. Constantly switching between these two modes — that’s my daily workflow.
Convergence and Divergence
I’ve organized my AI usage into two categories, eight actions:
Convergence: Collect, organize, analyze, summarize. Compressing scattered information into something structured.
Divergence: Brainstorm, create, imitate, develop. Expanding structured material into new output.
Here’s a concrete example. Say I want to build an analysis tool. The process looks like this:
- Converge: Feed sample data to the AI along with my own independent analysis. Ask it to organize my key points and cross-check the data to tell me what I missed.
- Diverge: Have it brainstorm how to evaluate the gaps, what additional suggestions could be incorporated.
- Converge: Integrate the new suggestions and distill them into an analysis Guidebook.
- Diverge: Use the spirit of the Guidebook to craft prompts that make the AI analyze a new dataset the way I would. Meanwhile, I also analyze it independently.
- Converge: Compare the AI’s output against mine, then calibrate the Guidebook.
- Diverge: Finally, turn the Guidebook into an automated program or tool.
Throughout this process, you need to keep pushing back. If the output isn’t good enough, tell it exactly what’s wrong and let it correct course. If you’re not sure how to communicate what you want, you can also ask it: what steps do I need to go through to achieve my goal? Let it walk you through it step by step.
Your flexible thinking is what unlocks AI’s power as a strategist.
Personal Knowledge Base: The Maniac Recording Method
I now document how I do everything. The method is simple and a little insane: I talk to myself out loud while working and record it. In situations involving other people, I record with their permission. Then I transcribe all of it.
Within roughly ten samples, you can build a personal know-how knowledge base assistant. From that point on, your productivity skyrockets, because your AI is no longer a generic language model — it’s a clone that actually understands how you work.
I recommend using the Gemini API with a batch transcription script. Videos, articles, courses — transcribe everything first, then do a rough categorization by content. Once the categories look right, start distilling. Strip out the filler, keep only the essence. I started with 800,000 words across 350+ files. After distillation: 120,000 words, about 50 files.
The last step is going back to compare against the originals to make sure nothing important got lost in the distillation. You can’t skip this.
Running Multiple Agents in Parallel
Four terminal windows running simultaneously is now my normal. One handling video migration, one doing video translation, one running website SEO optimization, and the fourth running personal finance simulations. Each window might have multiple sub-agents running in parallel on top of that.
The $200/month Claude subscription? In terms of felt value, I’m getting $2,000 or more out of it.
But I’ll also say this: most people genuinely don’t need the most extreme setup. Learn to manage your agents well during your focused work hours first. Learning to manage AI effectively is the real priority. When you’re away from the computer, let yourself actually rest. Right?
The Line Between Work and Life
With AI in the picture, the boundary between work and life gets blurry. You can interact with AI at any moment, and the pressure builds. It’s like living right next door to the office — most convenient, sure, but we still need a commute as a buffer of time and distance to maintain work-life balance.
What I do personally: I let AI agents work overtime for me until 5 AM, but I’m in bed by 11 PM. The time when the AI is running and I’m not idle, I spend thinking about how to evaluate the output, how to give clearer instructions, trying to understand the logic behind its changes. But when it’s time to rest, I rest.
What to hand off to AI, what to keep for yourself — the entire workflow needs constant, dynamic restructuring. This isn’t a one-time decision. It’s a judgment call you make every single day.