How much work can one person + AI really handle?

Many people think the key is to spin up multiple agents and run them in parallel. The truth is, what most often determines how much you can deliver each day is a few more fundamental things.

Task cards
For work that takes more than 10 minutes, I don’t just hand it to AI and start. Instead, I spend a few minutes writing a lightweight card first—clearly stating the goal, boundaries, acceptance criteria, and at most how many times it may be tried. The AI only counts as done if it meets the card’s standards, not if it “feels close enough.” The card is saved as a JSON file. Close the chat window, switch models, come back the next day—you can just read it and continue.

Without this card, work easily gets half-done with hand-waving, or gets stuck forever in “just one more revision.”

Divide tasks across models
I don’t let the same model handle the whole thing end to end. Claude mainly handles judgment and overall control. Codex is used for cross-verification and technical review. Cursor implements engineering tasks based on specs that are already clearly stated. Grok focuses on writing the body text of public-facing content. Switching is done by stage, not re-selecting every round. It’s not about using whichever is cheaper—it’s about which model’s ability at that stage can drive quality to where it needs to be. Using the wrong model means the time spent rework later can be more than what you supposedly saved.

Write all status into files
Task progress, decision rationale, handoff notes—I almost never leave it only in chat history. Everything goes into the repository’s Markdown and JSON files as the single source of truth. Start a new session, or switch models directly—read those files first and you’ll know exactly what step you’re on. You can close the chat window anytime; the system state won’t be lost.

Relying on the chat window to keep notes will eventually break—because the window resets or the model changes.

Quality gate + data recovery loop
Before sending content out, it has to pass a quality gate. Only hard facts can be sent properly; anything that’s just speculation must be explicitly labeled as “a brainstorm.” After posting, I check data after 24 hours. Only if saves and real interactions cross the threshold do I continue—producing the English version, cutting videos, and archiving to the website. If it doesn’t hit the line, I stop immediately and don’t add more time.

The time saved can be used to work on the next truly wanted thing.

Once I fixed these four items, I found that a single person with AI delivers more stable output than when they used to open agents everywhere—and with far fewer unfinished endings.

When you work with AI, what few rules are mainly managing the work? Tell me your version.