Writing three articles for Mandarin Daily News: how it actually went

This summer, I published three articles in the education section of Mandarin Daily News on how teachers can bring generative AI into the classroom:

• 7/29 "Reorganize your lesson materials: use prompts to create new activities"
• 8/5 "Plan a neighborhood walk: let AI be your lesson-prep partner"
• 8/12 "Generate fake scams: build fraud awareness through games"

I've already written about what the three articles say in another post (The best use of AI in the classroom isn't answering questions. It's building practice grounds）。

This one is about the process.

From pitch to print, three parties had a hand in these articles: me, AI, and the editor at Mandarin Daily News. Going back through my folders, the drafts, revisions, and the versions the editor sent back are all still there. Laid side by side, it's quite clear who did what.

The bottom line first:AI preps, the editor tastes, I cook.

1. The pitch: five topics became four

I started by sending the editor five topics.

Her reply was quick and direct: the structure of the second half was risky to tackle at this stage, so she merged the first and third points into four topics. Her reasoning was that one article, including images,usually runs about 1,300 Chinese characters, and with four topics each getting their own piece, even one article risked feeling crowded.

This was the first constraint I hadn't thought of, and one AI would never have told me about: page space.

AI has no sense of page space. Ask it to write, and it writes until it feels it's done. Every first draft I produced afterward came in between 1,500 and 1,800 characters, and every one had to be cut.

2. The most important conversation: should I use a real textbook as an example?

I asked the editor: should I use a specific grade's curriculum as an example? Say, a second-grade textbook?

I still remember her answer:

Being specific does sharpen the focus, but our teacher readers are very conscientious, so being too specific can sometimes carry risks.

Most of our columnists are teachers. Authors who aren't teachers usually approach things at the level of concepts.

Generally, I don't use "how to teach a particular lesson" as a demonstration, because that gets into teaching practice, and some teachers may be a bit more sensitive about it. A theme-based approach is more flexible, and it avoids issues with established knowledge frameworks.

In plain terms:I'm not a teacher. Someone who isn't a teacher, telling teachers in a teachers' publication "here's how to teach this lesson," crosses a line.

That's knowledge of the audience. AI doesn't have it, and neither do I. Only someone who receives teachers' submissions and sees teacher readers' reactions every day does.

Same day, three versions

That afternoon I used AI to produce three versions of the first article. The timestamps are still there:

• Version A (13:01): Purely conceptual. "Take a narrative text with characters and a plot as an example." No lesson title, no actual output.
• Version B, "the compromise" (16:02): In between, I found a real lesson plan for a second-grade text, "Making Breakfast Together." Version B described the lesson's content (the family splitting up chores, an egg like the sun, tomato slices like wheels) without naming it.
• Version C, "the worked example" (16:30): Named the text directly, included the skits, riddles, and quiz questions AI actually generated, and tied in the textbook's existing "...looks like..." sentence-pattern drills and emotion cards.

AI's value here is obvious: three directions, all built in one afternoon and laid out on the table for comparison. In the past, writing three versions meant I'd probably only write one, and gamble.

But once the three versions were laid out, choosing one wasn't something AI could decide.

The final compromise

Here's how the published version handled it:

• The text is named, but framed as "'Making Breakfast Together,' a narrative text common in lower grades." It'ssource material, not a teaching demonstration.
• The section from Version C tying in sentence patterns and emotion cards was cut entirely. That section was the most "instructional," and the most likely to make teachers feel "you're telling me how to teach this lesson."
• Lots of practical detail was added: all three tools work, a three-step process, and every section comes with a prompt you can copy directly.

In other words:Teaching judgment stays with the teacher; I explain the tool clearly.

I'm not teaching you how to teach this lesson. I'm teaching you how to get AI to generate material for you. That's a position someone who isn't a teacher can stand on in a teachers' publication.

3. AI loves to moralize; the editor loves to cut it

Comparing the drafts and printed versions of all three articles, one pattern stands out.

Every one of my drafts ended with a line like this:

• "The tool offers possibilities; judgment still belongs to the teacher."
• "AI is just a tool; the value is created by teachers."
• "Judging which fall is meaningful for a child still belongs to the teacher."

The sentences are polished, and the point is right. But say it in all three articles and it's preaching.

In the printed versions, most of these lines are gone, or shrunk to a short practical reminder, such as "AI may come across as too realistic or too fake; teachers should adjust the intensity to suit their students."

My understanding now:Telling teachers "the judgment is yours" is a bit redundant in itself. Teachers already know that. What they want to know is whether it works, how to use it, and what to watch out for.

AI has a writing habit: it likes to tack a punchy one-liner onto the end of a paragraph. In one article it feels powerful; read three in a row and it's glaring. AI can't cut these on its own, because it thinks they're the point.

4. Words the editor changed

The editor didn't just change paragraphs. She changed many individual words:

Some follow the newspaper's style guide (臺, 規畫); others are about tone ("kids" is a parent's perspective, "students" is a teacher's).

Changing [Prompt] to [Say to AI] really won me over. "Prompt" sounds like programming; "say to AI" is just talking. For a teacher who's never used AI, the latter is a much lower barrier.

The headlines were all rewritten too. Mine read like magazine titles; the editor's were newspaper headlines: you can tell at a glance what the piece is about.

By the revised draft of the third article, I'd already made some of these changes myself (kids → students, teachers → educators). I was learning the editor's rules, not AI's.

5. The second article: the editor asks questions in the draft

In the second article, "Neighborhood walk," the file the editor sent back had her comments written in parentheses right in the middle of my paragraphs. Five in total, and each one changed the article:

1. "Using AI (which one?)" I wrote "using AI," and she asked which one. If teachers are going to follow along, they need to know which website to open. From then on, every article specified "Gemini, ChatGPT, or Claude all work."

2. "This is where you should mention the core concept of what AI says, and fact-checking the information AI provides." I'd only included the list of neighborhood features AI gave me, not whether the list was accurate. The editor pointed out that teachers need to be able to verify local information AI compiles. The revision added a section: ask AI to cite a source for each item, check them online, and have the teacher review the sources before making the final call.

3. "AI says sixty years old. That's an awkward age. Should you use the idea of being older than grandparents?" I'd originally written that the old camphor trees were "older than most kids' grandfathers." The tree age was AI's estimate and not necessarily accurate, and comparing it to students' family members was odd anyway. It became: "This row of camphor trees was planted around when the neighborhood was built. The trees' age is the neighborhood's age." No need to cite an uncertain number, and it naturally brings in the neighborhood's history.

4. This was the longest comment, and the most important:

Since there's such a clear "Who gave the sky back to Minsheng Community," who is this walk supposed to find? How do they find them? Why did this person want to give the sky back to the community? As it's arranged now, the text reads more like AI planning a stroll, and it's hard to find where the lesson planning comes in. ...Or could the walk be an exploration game?

AI had given the route a great name, "Who gave the sky back to Minsheng Community," and I copied it straight into the article without noticing that the name itself was a puzzle. The whole piece was about "AI lines up three stops for you," and it read like a stroll.

One question from the editor, "Who is this walk supposed to find?", and the whole piece was rewritten: students become little detectives, each of the three stops gives a clue (someone left open space, someone planted trees along the street, someone buried the power lines underground), and a final "case-solving discussion" reveals the answer.

That structure wasn't AI's idea, and it wasn't mine. AI came up with the name; the editor saw the lesson inside it.

5. "Some of your AI responses actually answer the questions above. I suggest putting them into the text, because once the images are laid out on the page they can only be illustrative. Readers won't be able to see that text." I'd included several screenshots of AI conversations, assuming readers would get it from the images. But in a newspaper, images shrink to a single column width, and the text is unreadable. Anything important has to go in the body.

This is the easiest thing for a web person to forget: print has no "click to enlarge."

6. The third article: the editor edits directly

The third article, "Safety education," went through the most rounds:

1. First draft: About 1,800 characters.
2. My revision: Cut to about 1,500 characters, wording shifted to a teacher's perspective, and added "To check whether the scam scripts are realistic, compare them with the National Police Agency's 165 Anti-Fraud website."
3. The editor's revision: The editor edited the file directly.
4. My final pass: Smoothing out the editor's version.

In step 3, the editor did three things I hadn't thought of:

• Moved the "teacher oversight" section from the end of the article to the beginning. With safety education content, warnings have to come first. You can't wait until readers have finished reading how the game works.
• Added a suggestion from the classroom: before designing, observe which games and videos students are interested in, and switch to topics that resonate. That's something only a teacher would think of.
• Softened "teachers must run through this step themselves first" to "teachers can adjust the intensity themselves." "Must" is a command, which steps right back over the line from section 2.

Some of the sentences the editor inserted directly didn't flow yet. For example, a suggestion ended up inside the quotation marks of a prize-winning message, or a sentence stopped halfway.

At that point, I did just one thing:Smooth the sentences, without changing a word of the direction. If the editor wanted the warning up front, it went up front. If she wanted the point about observing interests, it went in. I only fixed the grammar.

This is the step in collaboration where things most easily go wrong: the author gets the draft back, sees it's been changed, and the first instinct is to change it back. But what the editor changes is usually the part you can't see.

7. Where AI genuinely does better than people

After all this about what AI can't do, it's only fair to cover what it can.

First, it builds the demos outright. The quiz game and the anti-fraud challenge game mentioned in the articles are real web pages AI generated from the prompts in the text, put online and turned into QR codes. The printed versions included game screenshots and a "click here to try it" link. Readers can play them right after reading, proof that this really works and isn't just theory.

Second, comparing versions. Three versions (A, B, C) of the same direction, the same article cut into a short and a full version: all laid out in one evening for me to choose from.

Third, the map and route for the second article. For the neighborhood walk piece, AI inventoried the neighborhood's features, linked them into a route, and marked the map. The screenshots in the article are of the actual conversation. (The article also notes where it got things wrong: it marked the second stop at a shop entrance, when the teaching focus was the old camphor trees along the street. That passage itself demonstrates "AI drafts, people revise.")

Looking back

If it had been just me and AI, these would have been three smooth, long articles that all ended with something wise. And teacher readers would have thought: another outsider telling us how to teach.

What the editor added were four things neither AI nor I had:

1. Page space: 1,300 characters, no room for even one more topic.
2. Readers: Teacher readers are conscientious; authors who aren't teachers should stick to concepts, not teach lessons.
3. Tone: "Students," not "kids"; "say to AI," not "prompt"; warnings go first.
4. A sense of curriculum: AI laid out a route; the editor saw a lesson. A question like "Who is this walk supposed to find?" can only come from someone who carries a classroom in their head.

AI let me try more; the editor showed me where to aim and when to stop.

What I did myself was turn one remark from the editor into a full set of revisions.

When she said "some teachers may be a bit more sensitive," I had to decide exactly how to revise: find a real lesson plan, have AI write three versions, pick the path of "keep the lesson title but treat it only as material," cut the sentence-pattern drills entirely, and add prompts readers could copy directly.

When she asked "Who is this walk supposed to find?", I had to tear the whole piece down and rebuild it as three stops, three clues, and a final case-solving discussion.

I ran every prompt in the articles myself first. Every game was actually built, put online, and printed as a QR code.

AI preps, the editor tastes, and I'm the one cooking. AI preps fast, chopping three versions in an afternoon; the editor has a precise palate, knowing in one bite where it's too salty or which dish went to the wrong table. But deciding what goes in the pan, how much heat, and whether the plate going out is fit to eat is the cook's job. For that part, no tool can do it for me yet.