People always ask me: “You publish every week, with a day job and small kids. How?”
I run Human+AI like a newsroom of one.
I’m a trained journalist, and the craft hasn’t changed: You find the story, research it, report it, get the angle right, verify everything, THEN you write it in your voice.
AI helps me with 3 jobs:
Deep research. It pulls sources faster than I can alone.
Contrarian. It argues against my take before a reader has to, and it does that more than once.
The unglamorous middle. The SEO pass. Turning a finished essay into a LinkedIn post and three Substack Notes. Rebuilding a table. Checking whether I already made this point in March.
A lot of AI writing advice doesn’t tell you that what makes the work good is what I don’t ask AI for. And there’s still a lot of work for me to do.
The 4 things that make Human+AI mine
1. The angle. Anyone can summarize the news. Deciding what to say about it, the piece only I would write, is my job. AI can’t do that for me or anyone.
2. Fact checking. I check every fact, stat, and claim against a primary source. AI hallucinates confidently and trust is my biggest asset. I’ve gotten things wrong, and when I do, I correct it fast and publicly.
I don’t verify to the depth a professional fact checker would. I rely on published research and reporting that’s already been through someone else’s desk, which is standard practice in many newsrooms.
3. The voice. AI can draft in a voice. It can’t write in mine, and that’s the part you subscribe for (I think!). There are all kinds of hacks out there for getting AI to write in your voice. So far, none of them have worked for me.
4. A real editor. My husband James is a former journalist, and he copy edits almost every piece. He gets his own step in my framework. (Being married to your editor may be the only real advantage of marrying a journalist. KIDDING!)
A lot of AI writing advice is needlessly snobbish. My take is that it’s okay to use AI to write. Especially if writing is a barrier to saying what you want to express but can’t, if you’re writing in a second language, or if you’re moving an idea quickly and style isn’t the point. AI literacy is a right, not a privilege, and that includes the right to use the tool without apologizing for it. Just verify what you’re publishing is true.
About that new label on my posts
Which brings me to the label that showed up on Substack last week. On July 21, Substack turned on AI detection, built by Pangram, which scans posts, notes, and comments, and tells readers whether a piece looks fully human-written, lightly AI-assisted, moderately AI-assisted, or fully AI-generated. Substack says they don’t want Substack to turn into LinkedIn. They don’t want readers spending attention (or money) on work no one thought through.
Pangram reads finished text and decides which cluster it falls into, human or machine. It learned by studying human writing from before 2021 alongside AI versions on the same topics.
It can’t see the writing process. But the process is the point.
A detector can’t tell you who chose the angle, whether a piece was fact-checked, or how much thinking the writer did. It reads the surface and returns a number that looks forensic, and the image of precision feels misleading to me.
When Stanford researchers tested a batch of GPT detectors in 2023, more than half of TOEFL essays written by non-native English speakers were wrongly flagged as AI-generated, while essays by American eighth graders were classified almost perfectly. People writing in a second language use a narrower band of vocabulary, and machines do too. Sam Illingworth from Slow AI writes that sometimes neurodivergent people are wrongly flagged, too. (I’m neurodivergent.)
Pangram says its false positive rate is low, including for non-native speakers, but those are company-reported numbers. Renée DiResta, who writes Agents of Influence, declined to audit them and noted she has hit false positives herself, as have plenty of other Substack writers.
I analyzed a piece of writing I did on Grow Here, which I write wholly, though I run it through Claude for edits. It said much of my original writing was “AI-assisted.” And it will sometimes wave through a machine-written post. Both mistakes are cheap for the detector and expensive for the writer.
All that to say that AI slop is real, but Pangram doesn’t seem to be the fix.
How I use AI: Steal this framework
Eight of the steps below are the craft, and a reporter in 1970 would recognize every one. For those 8, AI doesn’t change the work. It changes how long it takes.
Steps 6 and 10 didn’t exist 20 years ago. No one used to have to make their work findable by a machine, and no one was their own distribution team. But they are part of what it takes to publish these days.
The little ✋ icon means the step is 100% human led.
Scan. ✋ Read the news and the newsletters, listen to the podcasts, find the story worth telling.
Angle. ✋ Decide the take. Write a few, pick one. Pressure test with AI.
Report. Research to support and break the angle.
Draft. Outline, then spar with AI about my own argument. Draft, spar again. As many rounds as it takes to find the crux of the story.
Verify. ✋ Check every claim at the source. No exceptions. An old journalism adage applies here: “When in doubt, leave it out.”
Optimize. Make the finished piece findable by search engines and AI tools.
Voice. ✋ The rewrite and the final edit. Mine.
Edit. ✋ James (husband and editor) reads it and flags anything that doesn’t make sense. One in a while, I go back to Step 2.
Ship. Build the visuals, one last read on mobile, publish.
Distribute. Turn it into a LinkedIn post and a few Substack Notes. Mainly AI.
Five of the 10 are my brain. I use AI as a thought partner for some.
The habit to take from this today
If you do nothing else, before you draft anything, write your angle as a single sentence. Not the broad topic, your thoughts on the topic.
“AI agents are coming” is a topic. “AI agents won’t absorb the busywork they promise” is an angle.
If you can’t write that sentence, you don’t have a piece. That rule will do more for your writing than any tool on this list.
Why I’m telling you this
A newsletter about AI literacy that’s cagey about its own AI use doesn’t make sense. And no AI detector will tell you this. A detector reads the words and assigns a number. This post explains the process.
This is what good AI use looks like in a real job. That’s the framework, and it’s yours, for free.
If you want the detailed operating manual, including which AI tool does which job, the prompts I’d give you to start, and the process on one downloadable page, try a free paid subscription to Human+AI:
Here’s how I run each step. With tools, prompts, and judgement calls.
Step 1. Scan (No AI)
This is the one step where I use no AI at all. It’s also my favourite thing to do, which is thinking about how AI is changing the world.
I read. Newsletters, news, and a rotation of podcasts for my commute and laundry folding. I’m looking for stories that either tell you something important or teach you something useful, and I find that search tools aren’t good at finding that. Search finds what’s popular. I’m looking for what’s useful and under-covered, and that journalistic judgement only comes from having read a lot.
I listen to Hard Fork, I subscribe to many great Substacks including SheWritesAI, Code Like A Girl, Jenny Ouyang, Katie Harbath, I read books about AI, I watch YouTube tutorials, and I read the New York Times, The Globe and Mail, the CBC, the BBC, The Guardian, and more.
Step 2. Angle (My call, with AI’s help)
I write a few possible angles, then pick the one that’s most useful and least obvious. The first angle you think of is usually the one everyone else thought of, too.
I hand work to AI and ask it to challenge me and my assumptions. Asking a model to choose is a trap, and you will end up with the same story as everyone else.
Steal this prompt: “Here are three angles I’m considering. For each one, tell me who would disagree and why, what evidence would sink it, and whether it’s already been said. Don’t tell me which one to pick.”
Step 3. Report (Perplexity + ChatGPT/Gemini + Claude)
Perplexity for specifics and primary sources that back or break my angle. It’s a research assistant. I follow its cited sources back to the originals, and if I can’t trace an answer, I don’t use.
ChatGPT to stress-test. What’s the strongest argument against this? What am I missing?
Claude to think through the complicated parts, or turn messy notes into a clean outline. With my skills and folders attached. (I pay for this tool.)
Every output is a lead, not a fact.
Steal this prompt: “Here’s my angle. Make the strongest possible case that I’m wrong. Then tell me what a smart skeptic in my audience would ask first.”
Step 4. Draft (Claude + ChatGPT)
This is the longest step and the least linear. I outline, then argue with AI about my own outline. Then I write. Then I argue again. I go as many rounds as it takes to find the crux of the piece, which usually isn’t what I thought when I started. Somewhere in there I’ll talk to James and listen to which parts he pushes back on.
The rounds are the point. In the words of Hemingway, “The first draft of anything is shit.”
Steal this prompt: “Here’s my angle and my notes. Draft a first version in plain, direct language. No hype, no jargon, short sentences. This is a starting point I’ll heavily rewrite.”
Then, on every pass after that: “Where is this argument weakest? Just show me.”
Step 5. Verify (My call)
Every fact checked against a primary source. The non-negotiable: never publish a fact you got from an AI without checking it at the source.
Steal this prompt: “List every factual claim, statistic, and quote in this draft as a checklist, with the specific source each one needs.” Then I check each one myself.
Step 6. Optimize (Gemini)
A clickable headline is almost never the SEO friendly headline, but Substack lets you use both.
I hand the finished piece to Gemini and ask for an SEO / AEO read. A list comes back, and I take about half of it. The parts I always take:
The display headline and the SEO title. The headline on the page can stay Substack-y. The SEO title, which is what shows in search and on social, leads with search terms.
The meta description. 150 to 160 characters, with keywords in it. If you leave it blank, Google scrapes your opening paragraph, which was written to hook a human, not to summarize.
The slug. The URL that you can put into the SEO settings in Substack.
Put the main keyword in the first hundred words.
The AEO half is newer. The goal there is to be the thing an AI cites when someone asks it a question in your subject area. In practice, that means answering the question plainly and early instead of burying the answer under 800 words of setup, using question-shaped headings, and being specific enough to be quotable.
Steal this prompt: “Here’s my finished piece. Give me an SEO and AEO audit: an SEO title under 60 characters, a meta description of 150 to 160 characters, a keyword-led slug, keywords this piece could rank for, and which of my headings are invisible to a search engine. Then tell me what an AI answer engine would and wouldn’t be able to quote from this.”
Note the position of this step. It runs after the piece is written and before the final voice pass, not last. It means I absorb the useful suggestions and then rewrite them in my own words. Don’t over-SEO your voice away.
Step 7. Voice (My call)
The rewrite and the last edit are mine. I cut any machine writing, throat-clearing, or bland writing.
Steal this prompt: “Point out the 10 phrases here that sound most like AI wrote them. Don’t rewrite them. Just flag them.” Then I rewrite them in my voice.
Step 8. Edit (James)
I’m very privileged to have this step.
James is a former journalist and my editor. He reads the finished piece and raises anything that doesn’t make sense. The question he’s answering is whether a reader can follow the argument from one paragraph to the next.
Most weeks he finds at least one passage I thought was clear and isn’t.
AI can help with this. But an editor with standards can’t be substituted. My advice: your editor doesn’t have be a journalist. Ask someone to read your work and tell you what they don’t understand about it.
Step 9. Polish (Canva + Claude Design + Substack)
Canva and Claude Design for header images and graphics.
Step 10. Distribute (Claude)
The finished piece goes back to Claude and comes out as a LinkedIn post and a few Substack notes in one pass. Each Note has to stand completely alone, make a point, and not summarize the article.
Then I read and rewrite them. Usually one of the notes is good.
This process is the reason this newsletter is sustainable for me.
Steal this prompt: “Here’s my published piece. Pull 3 Substack notes from it. Each one stands completely alone, makes exactly one point, and does not summarize the article. Then write a LinkedIn post, hook in the first two lines, one practical takeaway. Plain language, short sentences, no hashtags in the Notes.”
Free download: The week on a page
Here’s a downloadable document with all the steps included. Try a step this week and tell me how it goes.
Thanks for being a paid subscriber!
Want to share the love? Give a gift subscription to Human+AI to a friend!
