I'd Rather Be Writing Podcast: My experience trying to write original, full-length human-sounding articles using Claude AI
Tom Johnson 10/25/23 - Episode Page - 38m - PDF Transcript
Hi, my name is Tom Johnson and you are watching, listening to I'd Rather Be
Writing.com. This is the podcast version of content that I publish. I'm
trying to... I'm trying to record more podcasts. So for the last couple years I've
been doing a lot of writing, mostly just text, and I've decided lately that I
should really try to convert more of this written content into audio content. Now
especially because like AI image generators are so... make it so
easy to create images, I've also put these... I put together like a very quick
slide deck for this. So anyway, I'm going to kind of go through this post. This
post titled My Experience, trying to write original full-length human-sounding
articles using Claude AI. And I'm not going to read the post because people in
the past have told me they hate that. And so I'm going to try to... I'm going to try
to just speak it. I'm going to talk through it. Alright, so the gist of the
post is that you can use tools like Claude AI or other AI tools to write
full articles. You're not limited to just fixing a paragraph here or making a
small update there. You can actually write full articles that sound very
human-like, but it's, you know, hard to pull off. And I tried to do it and I
thought it was going better than it actually was. In this post or in this
podcast, I'm going to talk about some of the challenges in writing full-length
content as well as my experiments and techniques and my strategies here. So
first, the question is why? Why even try to do this? I can write, I don't need a
computer non-human entity, an AI to do the writing for me. I'm a professional
writer, a professional technical writer. I've been writing on my blog since 2006.
So gosh, like, a long time. I don't have any trouble cranking out posts. Why would
I want to try to get an AI to write a post? Well, first, I love experimenting with
different techniques. It's sort of what makes it fun. But also, you know,
there's this idea or sentiment floating around that, like, hey, AI might
automate a lot of tech writing jobs or it might kind of reduce the value of
writing. And, you know, if AI can write some parts of content, why can't it
write more parts? Why can't it write full-length content? So in this post, I
am going to walk through the techniques that I think can get partway there. But
first, let me talk about some research as I was kind of looking into articles to
try to learn about how to write with AI to see what's been written, what's been
done and so on. First of all, it's almost impossible to get through wall after
wall of marketing. Hey, use our AI writer type tool, whether it's Jasper copy AI
or like a dozen others. It seems like that's the only thing I keep finding in
Google. But I did come across a couple of articles that were really worth kind
of learning about one is one is it's called navigating the jagged technological
frontier. And this image here really is trying to depict like a centaur, which
was a key metaphor in contrast to a cyborg. Apparently, the researchers gave
some people some tasks and so on. And some of the tasks were more AI capable,
like there were tasks that an AI could do in somewhere less sort of AI capable
tasks. And the consultants, these were the users, the consultants who did the
best were the ones who kind of were able to only use AI for certain parts of
the tasks. Whereas the other group, which were termed the cyborgs just kind of
used AI wholesale for everything. And so the point of the research was to say
that whether AI makes you productive or not really depends on whether you're
using AI for the right sort of task. And whether you're judicious enough to be
able to switch on the AI tools for those tasks where it's appropriate and then
switch off that AI mode for tasks where it's not. And it's not always clear
like when you should be using AI and when you shouldn't. But definitely this
idea of a centaur where you're, you know, centaurs like half man, half horse. And it
refers to this ability to delegate the tasks. Are you, I don't mean delegate in
terms of like having somebody else do it, but well, actually, maybe I do. But
delegate the AI to do something versus not delegating it to the AI. Another
article. Actually, this is a podcast that I really like. This is a, let's see, it's
on everyday AI. And the podcast episode was called Future Crafting. And the host
interviewed somebody named Brian Sykes, who I wasn't really aware of, but I think
he's kind of very well known in the creative ad design kind of domain. And
like the creative professional kind of space. And they use this metaphor of a
director. And their basic argument is that like, as you start to use some of these
AI tools, you shift from being the actor, who's sort of performing, you shift from
being the writer who's constructing each word to the director who's shaping the
scenes, who's telling telling the story and is directing the actors about what
they should say and how they should perform. And the larger idea is that like
you're going to clear up a lot of mental bandwidth so that you can perform this
director type role, rather than getting down in the weeds and doing the grunt
work of wordsmithing and really pushing out each of the sentences. They weren't
really talking so much about writing as like design. But I think a lot of it
applies. You know, if you think about all the energy and effort that goes into
shaping sentences that are smooth, that read well, that are direct, clear, easy to
understand and so on. It's a lot of work. This idea of being a director kind of
really appeals to me because with each post that I write on my blog, I spend a
lot of time just re reading it and proofreading it and reading it again. Each
time I try to let a day in between the edits to give it some space. And it's
just really tedious, you know, like trying to get it to read well and to spot
all those parts that jump out at me with more space requires a lot of
energy. So if I can somehow kind of offload that wordsmithing to AI and
instead focus more time and attention on the logic and the flow and this larger
narrative, that would be kind of a big win. Alright, so let me move on to what my
three biggest strategies are for using AI to write. Now, like I said, it's not
like I've cracked the code on this by any means. But I'm just sort of sharing
what I've tried and why. My first strategy is to prime the AI with accurate
information. Second one to go paragraph by paragraph. And then third to balance
personal voice with explanation. And let me jump into each of these with more
details. Okay, so this first strategy is to prime the AI with accurate
information. This is why I've specifically called out Claude AI because it
allows you to load it with a ton of input. You can put like 75,000 words
into it. And that's like between the input and the conversation limit and so on.
So the whole session. But if you're if you're writing about something that
requires explanation, and you want to make sure that like anything you're
explaining is going to be more accurate and not just made up hallucinated and so
on. By having a lot of articles that the AI can use as source material, really
make a big difference. So let's say you're writing an article about some topic,
and you want to be able to sort of access easy explanations about different
facets of the topic, you could go and find, I don't know, a dozen articles that
really contain a lot of the information that you would then later be summarizing
or paraphrasing or referring to. Just sort of load it into the AI as a kind of
training right in the chat. And then later when you need to make those
summaries, it's going to be a lot more accurate. So I'm a real big fan of the
of the Claude AI precisely because it has such a large input source. Alright, my
second strategy here is to go paragraph by paragraph. A lot of people start out
and they think, Oh, I'm going to have, I'm going to have this AI tool write a whole
like post a whole article in one go. And when you try that, you realize that it's
terrible. It's like, it's usually way off the mark. But just to take a lesson from
software methodology. Remember how with Agile, people basically do a couple of
weeks of software development, and then they show the feature to a user to
sort of check in. And then the user feedback might have suggestions or
corrections and so on. And so you'll you'll course correct and go back another
couple weeks, make some more features, show it to the user, you know, and get
feedback and keep doing that. That's the whole idea of modern software
development with scrum and other sort of agile workflows. The previous method
method was waterfall, where you would just like spec it out, the whole project
out from the beginning, and then build it for two years or so. And by the time
you've finished building it, you show it to the customer, and they're like, wait
a minute, that's not what I wanted. But you've already gone so far in the wrong
direction, that it's sort of unrecoverable. Well, going paragraph by paragraph is
precisely the strategy that aligns with Agile, and it results in much better
output. And I'll jump back into paragraph by paragraph more a bit later. Alright,
my third big strategy is to balance personal voice with explanation. I actually
did an MFA in literary nonfiction at Columbia University way back when like
back in the early 2000s. And my biggest takeaway is probably this this balance
creative nonfiction, like a personal essay or another any kind of, you know, more
literary essay really gets a lot of its appeal from this balance of personal
experience, anecdotes, insights, coupled with explanation, and this the sort of
interweaving of the two is what makes it come alive. It also has kind of another
benefit. If you're doing a lot of the explanation through an AI tool, having the
AI write that explanation, and then you switch into the personal voice and the
anecdote, that personal voice is going to better disguise is probably the wrong
word, but will disguise the AI written parts more. And when you read something
that has like a an eye of first person, it sort of tricks the brain into just
assuming and believing that that all the parts of the essay are are coming from
the author, better communicate, you better connect with with the reader this
way. Okay, so those are my my three big strategies I was trying. And now let's
let me break down kind of this 10 step process so that I can more thoroughly
kind of go through my experiment here. All right, the first step is to define an
information pattern. This is kind of one of my favorite parts because a lot of
people don't even realize that there's like information patterns to things.
Certainly, if you're a tech writer, you're you're accustomed to information
patterns in documentation, a task has a certain sort of order. You've got an
intro with prerequisites and the problem that you're solving and you've got
steps and examples and so on. And with a with a blog post, like sort of a long
form blog post, or not even long form, but just something that's more of a
personal essay ish style. There's a there's a specific pattern, you sort of
have this hook that explains the relevance. You define some kind of issue
that that's presented, you ask a question of some sort, what are you trying to
answer? What is what is this thing you're wrestling with? You summarize prior
research, and maybe you critique some limitations and so on, you do some sort
of experiment, you're trying things out. And over the course of trying things
out, you tend to have some sort of epiphany that changes your perspective.
And then that leads you to the conclusion. I mean, it's essentially kind of
like the hero's journey, only instead of like a hero, you've got your on a quest
to figure out an idea or get an answer to something. Michelle de Montaigne's
original sense of the word essay was like trying out an idea or making an
attempt with some thesis or idea. So anyway, that's the information pattern
that I think works well. And you can even see it in in this sort of podcast,
this post. Alright, so let's move on to the second one, create an outline. Again,
like because we're going to go paragraph by paragraph, you have to have a sense
of like what the article is going to say. I tend to have a rough outline so that
I can allow for more flexibility in the creative process. I don't want to be
really rigid and locked down, but I definitely have a bunch of points and I
put those points into that information pattern that I was describing. Alright,
step three, let Claude know the context. Right. So when you when you first jump
into this AI session, I say something like this, you're going to write you're
going to help me write an article for my blog, I'm going to lead you paragraph by
paragraph describing what I want you to write, I will be the director and you
will be the writer, you will articulate my ideas in readable grammatical ways,
adopting a plain speaking direct style. If at any time my ideas are ill
conceived, you will push back and recommend better approaches. Are you
ready to begin? So that kind of lets Claude know, hey, we're going paragraph by
paragraph, you're going to do the writing, I'm going to whisper ideas into your
ear, you're going to kind of be my muse and so on. One little tip, don't say the
word essay, because this sort of triggers student essays. And a lot of times, if you
do that, Claude might say something like, hey, it can't really help you with that,
but it seems fine ghost writing blog articles. So anyway, you might run into
that. Okay, step four, calibrate Claude's language. You'll start out, I'll start out
by just describing my first paragraph. Remember this, you know, relevance hook
sort of part. But really, in this first paragraph or this first response from
Claude, you want to make sure that Claude is using the right sort of diction and
style. If it's too like, highfalutin, or whatever, if it's too grandiose or like
pseudo literary, you can try to correct, correct Claude and say, hey, simplify your
language more. And if it than if it like oversimplifies, you can say, hey, try to
strike a balance between the two and so on. You can go back and forth a little bit
until until it's sort of hit a style that you find kind of normal. Usually by
default, Claude is pretty like conversational and readable. But anytime I
try to say, hey, adopt a style of a New Yorker talk of the town type writer or
something, it always goes poorly. It sounds terrible, like somebody's got a
Saras in their, in their pocket and every sentence they're trying to use it and so
on. So anyway, after calibrating, move on to the next step here is to continue
paragraph by paragraph. Just I go through my outline, and I describe what I want it
to say in each paragraph, and then it writes it. Now, this is a part you may
say, well, geez, if you have to explain each paragraph, like that's half the
work or writing. Well, yeah, you have a good point there. And for sure, this isn't
really saving a ton of time. So this might be where the experiment kind of
fails. You know, maybe it saves only like 20% of time and isn't really worth it
because then you have to go back and maybe rewrite a lot of these paragraphs.
But anyway, this was my approach and my attempt. And I go paragraph by paragraph.
And it is kind of interesting because you can read how like you can read almost
in real time, the articulation of these ideas and more full fledged paragraphs.
And you can see what you're, what you're writing and the ability to sort of see
what you're saying gives you more distance to better evaluate and assess it.
So there's value in that. I don't quite, I'm not quite sure how much value
because there's also a lot of value in writing something yourself and sort of
going more slowly because that lets ideas form and solidify more. So I'm sort of
mixed here. But anyway, that was, this has been my, my attempt. All right, the next
step is to compile all the paragraphs into like into an article. Right. Some of
the paragraphs might have been written several times by Claude, maybe they were
in the wrong order and so on. Step seven is to edit the whole. And this is kind of
the fun part, right? Because you've reached a milestone and you're like, Oh, I've now
got it first drafts, first drafts are usually things that require a lot of edits and so on.
But edit the whole, see if it's got the right shape and feel and so on. And then you move
on to fine tuning the article. Now, at this point, you know, since Claude or whoever,
the AI has written a lot of this stuff. There might be a lot of like cliche phrasings or
phrasings that are unnatural or sound weird. So it requires quite a bit of fine tuning.
All right, step nine is to maybe incorporate some auto generated art.
One really interesting technique is to just like copy and paste a whole section
into Claude and say, Hey, generate or create a prompt for an AI image generator based on
the ideas here, and then just paste it into chat GPT plus with Dolly as the mode. It's
available in the plus version and it generates art. That's how I generated like all of the
slide art here. I did this very quickly and just sort of picked the most, I don't know,
one of the four images that it generated that I liked. So you can move really fast
if you want. All right, step 10, let the content sit a while, even though it was probably
pretty easy to get a first draft. And maybe you think, Oh man, I could create a whole essay
every day like this. You don't want to do that. You want to let things sit because in the back
of your mind, maybe you're allowing things to surface and you're realizing that, Hey, I really
didn't want that part or this is bugging me. That sort of like back of the mind percolation
takes a couple days, at least for me. All right, now how did these posts turn out? I've only written
like a couple of posts using this method. So like I said, this is this is an experiment.
And one of these posts was a the attacks as post and the other was embracing professional
redefinition. And I got some good feedback on both of them. For the most part, people like
the posts, people didn't come out and say, Hey, this sound, you know, this is obviously AI written,
Tom, why are you doing? No. So it wasn't bad. I didn't feel like they were great posts, but I
felt like they were decent. But one reader actually reached out to me and he's like, he said, you
know what, your latest posts feel a bit off. And he was trying to put his finger on exactly why.
It's really interesting. He let me let me just quote a little bit. He allowed me to quote him.
He said, forgive me if I'm out of line here. I've always looked to you, you know, as a thought
leader and pragmatist and so on. But lately, something just feels off. The writing doesn't seem
to be as refined or tight. Chunks are longer and less less Tom packed, using a kind of
interesting adjective there. They're less Tom packed, as I used to say, to refer to your uncanny
ability to really stuff a paragraph with value with an economy of well picked words. And he
specifically caught out the Deattaxus article. He said, the summary appeared especially generated
didn't have key key conclusions that I found peppered throughout the body. Still good stuff, but
just not like he's used to reading from me. He says AI tends to do this, it makes a nice
bulleted lists that don't really reflect the same scrutiny that a writer would apply before
figuring out which of the points are really relevant to an argument and so on. So this got
me thinking, because I was really like, I was really feeling pretty proud of this process,
thinking, man, I've totally nailed it. And then I see this reader's feedback and I'm like,
you know, I think he might be onto something. But it was kind of hard to sort out exactly what he
was saying. Why was the writing off? And this is this taps into really a larger question a lot
of people have had. They're like, you know, the the AI tools, they write well, but something's weird
about it. It's like something is off and different people say different things. I've read one article
where somebody said, it's the imperfections that we like, you know, the AI writing is too perfect.
Well, I did some more research. And I've got several ideas here. My first reaction was,
oh, maybe it's the uncanny valley. This is the idea that as you get closer and closer to something
that looks almost human, you cross over this sort of threshold where it's clearly not a robot and
you're like getting very close to human. And you hit a point where people actually start to feel
a lot more unease and revulsion about it. This image right here, I think is a good a good depiction
of the uncanny valley, because it's like kind of disturbing, because it looks mostly human, but
it's clearly not quite. Anyway, so I thought maybe maybe the writing seemed to be humanish, but
wasn't quite. And I thought, well, what if, you know, especially if somebody's familiar with how
AI writing looks, if you start to if you start to sort of smell that it's AI written, maybe there's a
inner revulsion at feeling that you're sort of being tricked, like, hey, I thought I was reading
a human. Instead, this is a machine. And maybe that leads to a sense of sort of betrayal and
distrust that then provokes a stronger reaction. I'm not really sure. But when I came back to the
reader's comments, he wasn't really saying that it sounded almost like Tom. He was saying something
different. He was saying that it wasn't really like there wasn't tight logic. It didn't, it was
rambling, or it didn't have the writer's discerning compactness of ideas and word choice.
So I found another article. The second article here is
by Laura Hartenberger called What AI Teaches Us About Good Writing. And it's a little more vague.
She touches on a number of different things. Let me just read a passage. She says, as readers,
we need to feel like the writer is paying paying attention to us, trying to connect. Chat GPT
cannot build a real connection with its reader. It can only imitate one. Reading chat GPT's writing
feels uncanny because there's no driver at the wheel, no real connection being built. While the
machine can articulate, articulate stakes, it is indifferent to them. I think I mistyped that.
Anyway, machine can articulate late ideas and experiences. It's indifferent to them. It doesn't
care if we care. And somehow that diminishes its power. Its writing tends not to move us emotionally
at best. It evokes a sense of muted awe akin to watching a trained dog shake a hand. Hey, look
what it can do. So my takeaway from her article is that there's sort of a connection, an emotional
connection through lived experience and sharing of experiences and in this sort of meaning and
paying attention to the reader that comes about from writing that is sort of lost when you don't
have a human to human connection there. And she also says perhaps the time spent writing matters
as much as having written. And I really like that because for sure when you wrestled with an idea,
and I've written many posts like on my blog for the last, you know, 20 years, those posts where
I'm really trying to sort out in my head what I think about something and wrestle with that.
They're a lot more meaningful. I mean, if you just have an AI write something, it's not as meaningful.
And in a lot of ways, you're short changing yourself from the writing writing experience. If
the value of writing is having written, then if you haven't written, then you're losing the value
of writing. And if you're doing that, if you've lost the value of writing, but you're still producing
words, it's going to be a hollow experience. It's not worth repeating. So anyway, very good
perspective and something to keep in mind. Let me look at one more article. Actually,
this is not even an article. This is my own little hypothesis here. I think that these tools are very
over agreeable and probably inevitably so. But I found that they sort of carry out your will.
In my self redefinition post, I initially started out with a comparison to one of Nietzsche's
Overman or Ubermench ideas and so on, because it was about like, hey, creating new values and
meaning and so on. But there's kind of some negative baggage with that idea. And I was probably
better off just not even steering into philosophy there. But Claude totally like went that way
because I was telling it to and it didn't push back. And I think a lot of these tools will
not really push back. They're sort of programmed to be very polite and collaborative and constructive.
And so on. I got into a fun little exchange with Claude because I was like, hey, you know,
a true friend would really push back. And it said, well, I definitely will if there's like
factually false information such as saying two plus two isn't is five or something. And I responded,
well, in certain systems where you have the sum being more than the individual parts and so on
with dynamic feedback loops, it's very possible that two plus two could equal five. And it was like,
oh, yes, of course, you're right. In that situation, it very well could be I'm so sorry.
You know, I should have been more more careful in my logic. And I'm like,
this is you're demonstrating the exact point you agree with everything I want to say.
And I think this over agreeableness could lead to a lack of having a crisp argument. If every time
you rely on AI, it sends back kind of this watered down explanation that doesn't
take a side in any kind of issue. Then that might reflect in the writing as well,
where I don't really have a crisp point. I'm not taking a side. And and the writing might not be
compact, like the reader was saying. This is just a hypothesis. Obviously,
you can use the tools in many different ways, but definitely be on the lookout for over agreeableness
and realize that the tools aren't going to push back and say,
are you really sure you want to reference Nietzsche? I mean, you're writing to
a technical communication audience, not a philosophy audience. And by the way,
you don't really know that much about philosophy either. It's never going to be that direct.
So let's just wrap this up. I think based on my experiment,
it would probably be best to try to use AI mostly for summary and explanation.
And then for the personal experience, more of the first person parts to not rely on it.
It's my sort of guess or hypothesis that by just using it more for summary and explanation,
you'll be more apt to have a more direct and clear argument if argument is sort of your game here in
the post. You're more apt to follow that argument and have more of a point and an opinion than if
you are relying on AI for that part as well. I do think though that using AI for summary and
explanation really makes things a lot easier. These tools are great at that and it can reduce a lot
of the cognitive overload for those parts. And that balance, again, coming back to the strategy
I was describing earlier, this balance between personal experience and summary and explanation,
they go back and forth, they interweave. I think it really makes for compelling engaging content,
at least for a blog or other online stuff. That really is a nice winning formula.
So like I said, let me just caveat this. I've only tried this for a couple of posts.
I have some news summary posts where I'll summarize some news and so on, but it's not the same as
trying to write like a personal essay. And I may decide this whole technique is not worth doing.
Maybe it doesn't really save much time and it's not worth it. Maybe I just have to rewrite everything
that Claude wrote myself because I want it to sound like me and have my ideas and phrasings.
And I could just cut out that whole rewriting process and be much more efficient.
But I also kind of think this might take a little bit of getting used to. It may be
something that I get better at and maybe I can fine-tune a process. Ultimately, I do think
that it could be useful in the workplace for writing documentation as well,
which is a totally different genre. It doesn't have personal experience or argument or voice,
three huge components, but it does have a high degree of accuracy. I'm talking about the genre,
high degree of accuracy and precision that also can be a challenge. So I'll hit that in another
post. If you have any questions or feedback, shoot me an email at tomjohtatgmail.com or the
contact form in my blog. I'd rather be writing .com. You can subscribe to my podcast pretty much
anywhere. Pocketcasts, iTunes, Spotify, you'll find it pretty much everywhere. And let me know if
you like this format with the slides. This is now on YouTube and so on. I needed a visual element.
It also sort of helps to talk through the points rather than relying on written text.
But it would be great to get feedback. And if you feel like reviewing the podcast, great.
If you don't, don't worry about it. But thanks for listening and see you online.
Machine-generated transcript that may contain inaccuracies.
You can use AI tools like Claude to help you write full-length content. By going paragraph-by-paragraph, you can direct the AI while seemingly maintaining your own voice and ideas. However, despite my attempts to use AI with writing, I've found that it's harder to pull off than I thought. I can come kind of close, but due to the way AI tools are trained, they inevitably steer into explanation more than argument. This can remove much of the interest from a personal essay.