独立开发者的一天：AI 真正改变了什么

I’ve read a lot of “AI transformed my workflow” pieces. Most of them describe the tools, not the day. Here’s the day.

8:30 AM — Reading Agent Reports

The morning used to start with email. Now it starts with agent daily reports.

I run a few automated agents that work overnight — monitoring, summarizing, triaging. By morning there’s a digest waiting: what happened, what needs attention, what was handled automatically.

It’s a different ritual than reading email. Email is reactive — someone wants something from me. The agent reports are more like taking over a shift: here’s the state of the system, here’s what changed, here’s what I need to decide.

The mental model shift is real. I’m less reactive in the morning than I used to be.

9:00 AM — First Coding Block

The role has shifted. I’m not primarily writing code anymore — I’m reviewing it.

I describe what I want, Claude Code produces a draft implementation, I read it carefully and decide what’s right. Sometimes I accept it almost verbatim. Sometimes I throw it out and try a different description. Sometimes the right move is to write it myself, because the thing I’m trying to express is subtle enough that the description would take longer than the code.

What hasn’t changed: I still need to understand the code. Reviewing generated code requires the same mental model as writing it. You can’t just run it and see if it works — well, you can, but that’s how you accumulate technical debt you don’t understand.

The output is faster. The understanding is still required.

10:30 AM — Stuck

TypeScript edge case. Forty minutes of debugging.

I ran it through Claude twice. Got reasonable-sounding explanations that didn’t actually solve the problem. Eventually figured it out by reading the TypeScript compiler output carefully and tracing the type inference chain myself.

This happens more than the productivity discourse suggests. AI is excellent at standard patterns. Edge cases — especially language-level edge cases that interact with your specific codebase — often require human debugging intuition. The kind you build by reading a lot of error messages over years.

The AI saved time elsewhere today. This forty minutes was all me.

12:00 PM — Information Consumption

I don’t browse feeds during the day. That changed about a year ago.

Instead: directed search. I have specific questions — what’s the current state of X, what did Y ship last week, is there a better approach to Z. I use Perplexity for these. Fifteen minutes, specific answers with citations, done.

Passive feed consumption never made me more informed in any durable way. Directed search does.

2:00 PM — Writing

Articles, documentation, internal notes. This is where the collaboration model is most visible.

My process: describe the structure and key arguments to Claude, get a draft, read it, rewrite heavily. The draft gives me something to react to, which is faster than starting from a blank document. But the voice, the specific opinions, the things I actually think — those aren’t in the draft.

Claude produces plausible text. It doesn’t produce my text. The editing process is where I figure out what I actually want to say, using the draft as a foil.

By the time an article is done, maybe 60% of the words are original. The structure is often Claude’s. The content is mine.

4:00 PM — Agent System Maintenance

This is the part that AI can’t help with much, and it’s instructive about why.

Maintaining agent systems — updating memory files, adjusting configurations, debugging why an agent made a wrong decision — requires understanding the full system context. The agents themselves don’t have that context about their own architecture. Neither does Claude when I describe the problem, because the relevant details live in a dozen files across a non-obvious file structure.

This work is slow. It requires me to hold the whole system in my head. AI tools help with isolated tasks within it, but the system-level thinking is mine.

Every complex system has this layer. AI handles well-defined subtasks. The architecture and the judgment calls are still human work.

6:30 PM — Pre-Shutdown Archive

I update daily.md files before shutting down. Log what happened, what decisions were made, what I learned.

This used to feel like overhead. Now it feels like direct payoff, because the logs feed back into the agent memory system. Tomorrow’s agents will know what happened today. The context doesn’t evaporate.

The habit formed because the benefit is immediate and visible. Good incentive design by accident.

What Actually Changed

AI compressed the execution layer. Implementation that used to take a day takes a few hours. Boilerplate is cheap. Standard patterns are cheap. The gap between “I have an idea” and “I have a working prototype” is much smaller.

The judgment layer is unchanged. What to build, why, for whom, in what priority order — none of that got easier. If anything, faster execution makes the judgment calls more consequential. You can pursue a wrong idea much faster now.

The unexpected side effect: more experiments. When execution is cheap, the threshold for “let’s try it” drops. I run experiments now that I would have skipped before because the cost wasn’t worth the information. Some of them turn out to matter. That’s new.

The day is faster. The thinking is the same. The ratio of thinking to execution shifted dramatically, and that’s the real change — not any specific tool.