AI Is Rewriting My Workflow, but Not Every Step Is Worth Handing Over to It

Lately, I've increasingly felt that AI is changing not just my coding speed, but my entire workflow.

Previously, my approach to a problem was fairly fixed.

See a requirement, think it through myself first;

After thinking, search for information;

After searching, break down the structure;

Then start writing code, documentation, and filling in details;

If stuck, continue searching, trying, and revising.

Now it's different.

Now, for many steps, I no longer default to "start from scratch myself."

I first think about one thing:

Is this step really suitable for AI?

This might seem like just adding a tool, but I increasingly feel it's actually rewriting how I work.

Previously, a task was more about "I do it all myself."

Now it's more like "I first judge which parts should be done by me, and which parts can be handed to AI to run first."

The difference is significant.

Because what AI truly brings is not just speed, but forcing me to redefine:

What must I do personally, and what is not worth starting from zero?

But I'm also increasingly certain of one thing:

Not every step is worth handing over to AI.

If I don't think this through, AI might not bring efficiency, but just another form of busyness.

---

Previously, my workflow was more like "pushing from zero to one"

In the past, I used to push through many things from scratch.

For example, when seeing a requirement, I would first walk through the logic in my head;

If building a feature, I would first break down the structure myself;

If writing an article, I would first think of my viewpoint, then expand bit by bit.

This approach has its benefits, of course.

It allows me to understand the problem more deeply and have more control over the structure.

Especially when making architecture decisions, trade-offs, and judgments, this "go through it myself first" process is actually very important.

But its problem is also obvious:

Many repetitive, low-leverage, draft-type tasks are not worth starting from zero every time.

Previously, I treated "doing the whole thing myself" as the default.

Now I increasingly feel that this default itself is expensive.

Not just in time, but also in attention.

---

What AI first rewrote was not my code, but my starting approach

I later realized that AI first changed not that a piece of code was written faster, but that my initial action when facing a task changed.

Previously, when I received a problem, the default was:

"I'll do it myself first."

Now, more often, my default becomes:

"Let AI run it first and see if it can complete the first round for me."

This "first round" is very important.

It's not necessarily the final answer, and often it's far from it.

But it can first help me complete things that are time-consuming but don't necessarily require me to start from zero.

For example:

• Help me list a structure first
• Help me turn a vague problem into a clearer one
• Help me provide several comparable options
• Help me generate a first draft
• Help me fill in mechanical but necessary parts

AI might not do these perfectly, but often it's enough to save me the first stretch of the journey.

And the first stretch is often the most energy-draining.

---

Now, these are the steps I most commonly hand over to AI

The main categories I now consistently hand to AI are the following.

1. Draft-type work

For example:

• Article drafts
• PR review prompt drafts
• Requirement breakdown drafts
• Document structure drafts
• First versions of SQL, scripts, configuration descriptions

Previously, I would often write a skeleton myself.

Now, I often let AI produce the first version, then decide how to revise.

Not because it gets it right in one go,

but because "having something to revise" is much lighter than "starting from a blank page."

2. Information organization and structuring

Some problems aren't hard because you don't know, but because the material is scattered.

For example, you have a bunch of points in your head but haven't organized them into a structure;

Or after discussing requirements, you have information but it's not clear enough;

Or you can think of several technical directions but haven't made a comparison table.

In such cases, AI is very suitable for the first round of organization.

Its most valuable aspect isn't necessarily "thinking for me,"

but "first flattening what I already have."

3. Repetitive but not entirely skippable work

Some things aren't difficult but always need to be done.

For example:

• Rewording
• Standardizing formats
• Adding documentation
• Generating basic code skeletons
• Converting expression styles

I could do these before, but they consumed patience.

Now, as long as I judge that they don't require core decision-making, I'm more willing to hand them to AI first.

---

But I'm also increasingly clear: some steps cannot be easily handed over

This is also the clearest point I've realized recently.

AI can participate in many tasks, but not all key steps are suitable for it.

Once you over-rely on it in certain areas, it might seem fast on the surface, but in reality, you're losing control.

1. Key decisions still need to be made by yourself

For example:

• Whether this requirement is worth doing
• What the trade-offs of this solution are
• Whether to prioritize product development or take on projects now
• What this architecture choice means for future maintenance
• Whether this feature is for validation or self-comfort

For such questions, AI can help list options, add perspectives, and remind me.

But the final decision, I increasingly won't hand over.

Because such decisions aren't just about information; they also relate to goals, resources, stage, and risk tolerance.

These things, models can assist with, but cannot replace.

2. The final structure and tone are best finalized by yourself

Especially when writing articles, I now have a growing feeling:

AI can help with the first draft, but whether "this piece ultimately sounds like me" still needs to be finalized by me.

Because AI easily makes articles "correct."

The structure is complete, the logic is smooth, and it even looks flawless.

But the problem is, it also easily makes articles sound inhuman.

More precisely, not like "me."

So now my requirement for AI writing isn't "you write it all for me,"

but "you first take care of the most laborious first step."

The subsequent judgment, trade-offs, deletions, and tone still need to be done by me.

3. Parts involving real business understanding cannot be skipped

AI is strong on "general problems."

But once it enters real business, real users, and real project boundaries, it easily starts to fill in gaps arbitrarily.

This is why I now don't fully hand over things that truly depend on context to AI.

For example:

• Real customer needs
• Project boundaries
• Product stage judgments
• Business reasons behind technical choices

On the surface, AI can also talk about these convincingly,

but if the context isn't fully provided, it easily gives you an answer that "looks reasonable but might not suit your current situation."

In such areas, if you don't check yourself, you can easily be led astray by AI.

---

The biggest change AI brings is not speed, but that I start reallocating attention

I now increasingly feel that AI's greatest value to me isn't "how many minutes saved," but that it forces me to rethink:

What is worth doing personally, and what is not.

Previously, I defaulted to doing many things from scratch.

Now I'm more like a "workflow scheduler."

I first look at the task, then the steps, and then decide:

• Which parts I do myself
• Which parts AI runs first
• Which parts I only review the results
• Which parts I must personally finalize

This change matters more to me than "code generation being a bit faster."

Because it affects not just one action, but the entire way of working.

Previously, I was more of an executor.

Now I increasingly need to allocate first, then judge, then finalize.

In a way, AI makes me more like managing my attention, rather than just managing tasks.

---

My current simple principle: use AI for high-repetition, low-judgment, and quickly verifiable steps

If I were to compress my current experience into one sentence, it would be:

High-repetition, low-judgment, and quickly verifiable steps are more suitable for AI first.

Conversely:

• High-risk
• High-judgment
• High-context-dependency
• High-personal-expression

These steps, I'll be more cautious about.

For example, when writing a WeChat article, I'm now more willing to let AI help with:

• Topic expansion
• Draft structure
• First version of the body

But I won't easily hand over the following completely:

• Whether this article is worth writing at all
• Whether the core viewpoint should be emphasized this strongly
• Which sentence sounds most like me
• Which paragraph should be deleted
• What image this article will reinforce when published

These things still need to be done by myself.

---

So for me now, AI isn't working for me, but rewriting how I divide work

This might be my most obvious feeling recently.

Many people understand AI as a more powerful tool.

But I now increasingly feel it's more like a partner that constantly forces you to re-divide work.

The question is no longer just:

"Can AI do this task?"

But:

"In this task, which parts are worth doing personally, and which parts are not?"

The difference between these two questions is significant.

The former easily turns AI into a flashy tool.

The latter truly starts rewriting the workflow.

So now I won't simply ask:

"Should I use AI?"

I more often ask:

"In this process, should this step really be done by me from scratch?"

Once you start asking this question often, many things change.

AI is rewriting my workflow—I'm increasingly certain of that.

But I'm equally certain:

Not every step is worth handing over to it.

Handing over some steps is efficiency.

Handing over some steps is losing control.

The difference between these two will likely become increasingly important.