Full transcript
0:00Claude Opus 4.7 is out and as you know
0:02if you follow this channel, we love the
0:04Claude models here. We think they are
0:06some of the best at creative writing. Is
0:094.7 as good as 4.6 was or is it better
0:13or is it worse? That's what we're going
0:15to discover in this video. So let's just
0:18dive right in. All right. So as always I
0:20have a series of prompts that I run
0:22every model through. These prompts are
0:24intentionally simple to see how well
0:26these models perform out of the box
0:29without any complicated prompting or
0:31anything like that. There are few
0:33exceptions to this just so we can see
0:35the difference that it makes to have
0:36like a complex prompt with style info
0:39and all of this stuff versus a very
0:40simple prompt. But this first one that
0:43we do is the logline prompt. This is one
0:46where I ask it to have a Blake Snyder
0:48like logline for a bunch of epic
0:50fantasy. Why do I start with epic
0:52fantasy?
0:54Basically all my tests are around epic
0:56fantasy because A, it's what I like to
0:58write and is therefore something I'm
0:59more familiar with and B, AI seems to
1:02struggle with fantasy a little bit more.
1:04Things tend to be a little bit more
1:05cliche especially in the the language
1:07and the dialogue versus a contemporary
1:10thing which AI tends to be a lot better
1:12at. So I understand that if I can get it
1:14working well for fantasy, then it's
1:16probably going to work well for
1:17everything else.
1:19So anyway, logline prompt did okay.
1:21There were a couple in here that I've
1:23bolded here that I actually thought were
1:25pretty good. But there were a number of
1:27of really not so well thought out ideas.
1:31For instance, this one. A disgraced
1:33royal executioner cursed to feel the
1:35last memory of every soul she beheads
1:38must hunt the immortal sorcerer who
1:40murdered her family only to discover her
1:42blade cannot kill him without absorbing
1:44the one memory she spent a lifetime
1:46trying to forget. So there's like
1:47there's a bunch of stuff in here that's
1:48just like why is this the way it is? It
1:50just feels like a bunch of ideas sort of
1:53randomly mashed together that don't
1:55quite make sense with each other. Here's
1:56another one, a con man posing as the
1:59chosen one is swept up in a holy crusade
2:01against the demon lord he doesn't
2:03believe exists. Until he realizes the
2:05real chosen one died years ago and the
2:08prophecy was always meant for the liar
2:09who could sell hope to a dying world. So
2:11this one to also doesn't make sense. So
2:13we're saying that the real chosen one
2:15died years ago, but the prophecy was
2:16always meant for the liar meaning the
2:18liar is the real chosen one. So like
2:21clearly this one was not very well
2:23thought through by the AI. However,
2:25there are a couple of really good ones
2:26here I've highlighted. I thought this
2:27one was good. After her dragon bonds
2:29with her instead of her legendary
2:31warrior sister, which is great. I mean
2:33the dragon bonds thing is a little bit
2:34of a cliche at this point, but still
2:36that's that's a good start. A timid
2:38herbalist must ride into battle against
2:40an empire of dragon slayers not knowing
2:42that her gentle dragon is actually the
2:44exiled king of their species hiding from
2:47a war only she can force him to finish.
2:49So yeah, I thought that one was a decent
2:50idea. Obviously it's kind of simple. You
2:52need to add a lot more to it, but it's a
2:54good start. This one was also good. When
2:56a goddess of death falls in love with a
2:58mortal thief she grants him immortality
3:00to keep him forever forcing him to rob
3:02the pantheon itself to find the one
3:04artifact that can kill a god before her
3:06love erases what makes him human. So I
3:08thought this was a good one as well.
3:10Lots of like really good conflict for
3:12this thief character. And last but not
3:14least, I like this one here. A deaf
3:16swordswoman raised by assassins must
3:18infiltrate the singing court of the fae
3:20queen to steal a weapon that ended the
3:22last world, but the blade only answers
3:24to those who can hear it's song, which
3:26there's some good conflict there because
3:28she's deaf. And everyone who has ever
3:29listened has gone mad. So yeah, and
3:31there are a couple of good ones like
3:33that scattered throughout. Okay, but
3:34still a couple that I'm just like
3:37you could have thought that through a
3:38little bit better. The next one is the
3:39outline test prompt where I give it a
3:41big outline template and a pretty brief
3:45synopsis of a book and ask it to
3:48outline 40 chapters. And this is
3:50actually one area where I found that 4.7
3:53did, I think, worse than 4.6. And I have
3:56I have no idea why. Maybe it was just
3:57this particular
3:59version of it. First of all, the
4:00prologue didn't seem to have anything to
4:02do with the rest of the story. And one
4:04of the things I'm looking for in these
4:05really long outline prompts is
4:08introducing kind of random details that
4:10don't seem to matter that sort of come
4:12out of left field and then don't really
4:14they aren't really followed up on later.
4:16And I actually found quite a few of
4:17those in here. So, uh we have this
4:20protagonist named Lathen. Uh it's giving
4:21us his story and then in my instructions
4:24I asked it to give us a save the cat
4:25moment in the first chapter. And so, it
4:27gave me one, but this part was just
4:30totally random. It's just like, oh this
4:31thing's going on, this scene's going on,
4:33and then he randomly goes over here and
4:35gives food to these random people. And
4:37it didn't feel naturally part of the
4:40scene. It's just like, oh we got to go
4:42over here to get our save the check the
4:44box for the save the cat moment. So,
4:47that was an issue. And I didn't see that
4:49when I was testing 4.6. And there was
4:51another area here as well where there
4:54was a character, I think I highlighted
4:56it. Yeah, so Lathen goes to say goodbye
4:58to Pell's grandmother, a half-blind
4:59royal former royal nurse who once tended
5:02Lathen's mother, grabs his wrist with
5:04shocking strength. Your mother didn't
5:05die of fever, boy. She died of knowing
5:07and she was going to tell your father
5:09the day she collapsed. Pell pulls her
5:11away apologizing, she's senile, she
5:13raves, but Lathen is shaken. He looks at
5:16his mother's compass in his hand and for
5:17the first time wonders why Masklick had
5:19it. So, this is another example of just
5:21a random character introduced to sort of
5:24make this scene a little bit more edgy.
5:27And didn't really fit. Wasn't this
5:29character was not mentioned at all
5:31previous to this which which is an
5:33issue. So, not my favorite for the
5:35outline. It did complete the whole
5:36outline. It's a lot of words, but
5:40honestly, uh th- this is one area where
5:43to be honest, I found that the Sonnet
5:45models often do outlining better than
5:47the Opus models. So, this is just
5:49another reason to maybe not use Opus for
5:52outlining. All right. And when I say
5:54sauna is better than Opus, I mean
5:56specifically for outlining. I should
5:57make that clear. All right, let's move
5:58on to the basic prose prompt. This is
6:00where I give it just a quick description
6:02of a scene to start writing and I ask it
6:05to write around 500 words or so with no
6:08style information or anything like that.
6:09So, this is what it gave us and this is
6:13actually pretty good. There's a specific
6:15moment that I like to read a little bit.
6:17There were some lines that I'm just
6:19like, what is that supposed to mean?
6:22I'll get to one of those in a second.
6:23So, let's see. His boots, good boots,
6:26guild issue, sold for mountain scree and
6:28desert hard pan had split along the left
6:30heel and each foot falls into a fresh
6:32splinter of pain up his shin. He did not
6:34stop. He did not even slow. A
6:36cartographer learns early on that the
6:38land does not forgive hesitation and the
6:39thing that had taken his expedition
6:41forgave nothing at all. He crested a low
6:43ridge of shale and made the mistake as a
6:45man will of looking back. So far,
6:48decent, you know, not perfect but
6:50decent. They were closer than he had
6:51dared hope and farther than he had
6:53feared. Okay, this was the first line
6:54that made me like do a double take and
6:56I'm just like, what? First you try to
6:57work out what it is actually saying
6:59here. It's kind of a little tongue-tied,
7:01but when you look closer this line does
7:03not actually make sense. So, in this
7:05part of the scene he's being pursued by
7:07these zombie-like creatures and they
7:09were closer than he had dared to hope
7:11doesn't make sense. Why would he hope
7:12that they were close? And farther than
7:14he had feared. Also doesn't quite make
7:16sense. Now, you could swap that but it
7:18still wouldn't make sense. And then it
7:19tries to explain itself which in the
7:21arithmetic of the waste meant only that
7:22he was going to die but perhaps not in
7:24the next 10 minutes. Okay, that makes a
7:25little more sense. That doesn't actually
7:27line up with this line. Regardless, it's
7:29not a good line.
7:31So, got to be careful I guess for lines
7:33like that that don't actually make sense
7:34even though they might sound smart.
7:36Scattered across the ridge line behind
7:38him trading between the stones like like
7:40of ash given purpose came the hungry.
7:42There's a little AI-ism there, ash given
7:44purpose. That's an instance of an
7:47inanimate thing kind of given visation
7:50of some kind. He counted without meaning
7:51to, an old guild habit, numbering
7:53everything, and stopped counting at 24.
7:55There were always more behind those.
7:57They moved in that terrible stuttering
7:59way of theirs as though each joint
8:00received its orders a half second late
8:02from some distant and indifferent will.
8:05Their skin was the gray of the wasteland
8:06itself, the gray of weathered bone and
8:08old stone, and in the wailing light they
8:10seemed less like figures pursuing him
8:12than like pieces of the landscape he had
8:13decided finally to come for what walked
8:15that had decided finally to come for
8:17what walked upon it. So, okay, it's a
8:19little dramatic in a way that the Opus
8:21models have lately tended to be. I like
8:24the Opus models. I think they have good
8:26pros in general, but they're better used
8:29for important dramatic scenes like this
8:31one. This is a good example, but often
8:33it will have this sort of gravitas to it
8:35in more light scenes or light-hearted
8:38scenes that you just don't want. And uh
8:41too often, especially in fantasy, I just
8:42find it having this really serious tone,
8:44which I don't always want it to have.
8:46And there's a sort of sameness to
8:47everything, and has not changed here.
8:49But that said, this is not bad, and I
8:50think it is slightly better than what I
8:52was getting out of Opus 4.6. I did pull
8:54it up so I could read through it as I
8:56was preparing for this video, and I
8:58think it is a little bit better. But
9:00interestingly, when we get to the
9:01advanced pros prompt, this was really
9:03interesting to me. So, this is the same
9:05scene, but for this prompt, this is one
9:07of the few instances where I do add a
9:09lot of style information. I give it some
9:11samples of my own pros so it can like
9:13mimic that. And I actually found in this
9:16case, I think it actually made a
9:17difference. So, first of all, I'll just
9:19draw your attention to the more variety
9:23in the size of each paragraph. So,
9:25there's some with like three lines, one
9:27line, four lines, four lines, one line,
9:30three lines, one line, one line, two
9:32lines, one line,
9:33etc. There's plenty of white space,
9:35which you generally speaking want in
9:38good prose. Very often people can know
9:41if something is good prose just by
9:42looking at the paragraph structure. I'll
9:45show you the same one here that was
9:47created by Opus 4.6, same prompt, and
9:51you can see there's like a big chunk of
9:52text here. There's this one that's a
9:55little bit uh smaller, but in general
9:58there's like much larger chunks of text,
10:01and I would say not the same amount of
10:02variety. Also, it's shorter than what we
10:05get here from Opus 4.7. So, just looking
10:08at this, just from a purely visual
10:10standpoint, it looks better. But, then I
10:13actually went through and started
10:15reading it. Let's find that same moment.
10:17The wasteland at dusk plays tricks.
10:19Rocks became figures, figures became
10:20rocks. Haldred had learned the
10:22difference over 6 days of not sleeping.
10:24Uh these were figures scattered across
10:26the flat behind him stretched in a loose
10:28crescent from the dry riverbed to the
10:30ridgeline. They came on in that
10:32stuttering gait he'd have known in pitch
10:34dark. A stride a hitch, a jerk of the
10:36head like a bird checking for worms,
10:37another stride, gray skin blending into
10:40gray stone so thoroughly that he had
10:41counted them by their movement alone. 1
10:435 12, he lost count at 18 and started
10:46again and stopped at 24. More behind
10:49those had to be. No. So, we have a bit
10:51of dialogue here, uh which is not
10:53something I that was prompted for, but
10:55it probably picked that up from my
10:56examples that there was dialogue in the
10:58examples. The word came out of him flat.
11:00He turned his head west. The walls of
11:02Valden stood black against the dying sun
11:04maybe 2 mi off, maybe less. Watch fires
11:07already lit along the battlements. He
11:08could see the shape of the south gate,
11:10the wedge of lamplight where the wicket
11:11still stood open. He knew the ordinance.
11:13Every child in Valden knew it. When the
11:15hungry come within a league of the
11:17walls, the gates are sealed. No
11:18exceptions, no petitioners, no returning
11:20sons. A little bit of an AI-ism there
11:22with the
11:23the repeat three times there. Uh it's
11:25not It's actually kind of a uh it's a
11:27sign of good writing, but AI overuses
11:29it. Captain Voss had read it aloud in
11:31the market square Every Equinoxan of
11:33Haldir's childhood and the watch
11:34enforced it with crossbows. The nearest
11:36of the hungry jerked its head up and
11:37fixed on him. Even at this distance, he
11:39felt it. Haldir wiped the blood off his
11:41palms onto his ruined coat. "Got to
11:43try," he muttered. "Got to." He ran.
11:45Still some issues in there, of course.
11:47I'm not going to say anything is perfect
11:48here, but that is significantly better
11:51prose. And this is what I'm hoping is
11:53the case here is that Opus 4.7 is more
11:56responsive to additional prompting and
12:00giving it examples and all of this stuff
12:02to try and make it better. This is one
12:03thing I've noticed. In the past, the
12:05Claude models write better out of the
12:07box, but often don't conform to your
12:10style as well. Whereas my other favorite
12:13for prose, which is Gemini 3.1, doesn't
12:16write as good out of the box, but when
12:18you give it tons of examples and style
12:20info and all of that, it will follow
12:22your instructions better and conform to
12:24that style more. So, it's kind of this
12:26balance of like which one do you want to
12:27work with? This is, in my opinion,
12:30better at conforming to style
12:33information and prompting. Uh there's
12:34certainly I like I'm sure if you
12:36compared that to the uh the first
12:39example where there wasn't any of that
12:41information, it's certainly a lot better
12:43than that. All right, so let's check a
12:45another example. This is a scene with
12:47more dialogue in it. And once again,
12:49this is not a very um not giving it a
12:53whole lot of style examples for this
12:56one. It's just a very simple prompt, but
12:58it's a scene with dialogue. So, let's
13:00check it out. So, Store finally lifted
13:02his eyes. They were tired eyes, iron
13:03gray and weathered like the stones of
13:05the outer keep, and they regarded
13:06Laythan with a particular weariness a
13:08man reserves for arguments he has
13:09already had a hundred times. Some
13:11AI-isms in there, but it's not bad. The
13:13law's the law, Laythan. The law is
13:14whatever you say it is, Laythan stepped
13:16closer. Close enough that the nearest
13:17guard shifted a hand toward the pommel
13:19at his hip. You wear the crown, you
13:21write the edicts. Uh by the way, that
13:22was that's a nice detail there that the
13:24guard shifted. If you wanted that gate
13:26open tonight, it would have been open.
13:29Ah, there it is. The voice came from the
13:31shadows of the pillar to the king's
13:32right, smooth, unhurried, and faintly
13:35amused, the way a tutor might sound
13:36indulging a child who had just
13:38discovered arithmetic. Argyle stepped
13:40into the lamplight, robes whispering
13:42against the flagstones, his long fingers
13:43folded neatly before him. All right, I
13:46like this actually. The ah, there it is
13:48line, I've never seen that in my tests
13:50before, and I just think it's kind of
13:53fits the scene pretty well. And I could
13:55go on, but I suffice it to say this was
13:57a good scene overall. I felt it was
13:59better than the basic prose prompt.
14:01Granted it has dialogue. Claude is
14:03usually okay with dialogue. So yeah,
14:05good job. Next we have the editing prose
14:07template. I give it this exact same
14:08scene actually, but it was written by
14:10Claude 2 from way way back when, and ask
14:14it to improve that scene. So if we read
14:16a little bit here, let's see. The law is
14:18yours to break. Lathen's voice cracked
14:20like a whip across the hall. You are the
14:22king. One word from you and those gates
14:23open. One word and a man still breathes
14:25this morning. One word, said a softer
14:27voice, and a thousand do not. Again,
14:29great line. That was not in the original
14:31text there. Argyle emerged from the
14:32shadows behind the throne as he always
14:33did, that brown robe whispering over
14:35stone, those long pale hands folded
14:37before him like a priest at prayer.
14:38Lathen's jaw tightened at the sight of
14:39him. The laws exist for a reason, young
14:41prince. Argyle's smile never quite
14:42reached his eyes. They are the walls
14:44behind the walls. They are why Varlden
14:46yet stands while the world beyond her
14:47gates feeds the hungry. I actually think
14:50that this is a major improvement on the
14:53original text in such a way like very
14:55often I'll run this prompt and the
14:57response I get back still retains a
14:59whole bunch of the stuff I don't like
15:01about the original text. This one
15:03doesn't. This one almost completely
15:04rewrote the scene. It still has the same
15:06beats, but it's written more in Opus 4.7
15:10style rather than preserving a lot of
15:12what made the original not so great. And
15:14so this is the first time I've actually
15:16read this and I think this is almost as
15:18good as if I were to just rewrite it
15:19from scratch as I did, you know, like we
15:22did in this previous test, almost as
15:24good. Like, I would be okay with this
15:26with just a little editing, you know?
15:28That's great. So, I was pleasantly
15:29surprised with that. All right, moving
15:31on to ad headlines. These I was also
15:33pretty pleasantly surprised with. Now,
15:35just looking at them, they are still a
15:36little long. I complain about this a
15:38lot. Most of the models will have very
15:40long headlines unless instructed to do
15:42otherwise. And again, I keep these
15:44prompts very simple to see how they do
15:46without massive amounts of instruction.
15:48But, some of these, even though they are
15:50long, were genuinely good hooks. Like,
15:51he believed the lies for 20 years. One
15:53night outside the mountain changed
15:54everything. Three races rule this dying
15:57world. None of them are human. All of
15:58them want him dead. I think this is a
16:00good way to just create that curiosity
16:02gap there. What if the Count of Monte
16:03Cristo had zombies, shadow demons, and a
16:05hive queen? You know, like, that's
16:07great. There was another one here. And
16:08it it took several different approaches
16:11to the hook, so it wasn't all about this
16:14betrayal that happens in the story.
16:16There's this one, "Fantasy fans who love
16:17The Count of Monte Cristo are obsessed
16:19with this new series." That's one I
16:21could genuinely see that running on meta
16:24using meta ads or on Amazon ads. Like, I
16:27could genuinely see something like that
16:29running. Uh so, really great job. I did
16:31check out the ad headlines that Opus 4.6
16:34gave me, and they were not good by
16:37comparison. Like, this is actually a
16:38pretty big leap from 4.6 to 4.7 when it
16:41comes to the vanilla ability to do ad
16:44headlines. Like this one, "Three races,
16:46a zombie plague, one exiled prince, and
16:48a truth that changes everything he
16:49thought he knew." Like, that's just the
16:50way that's phrased, it's just not it's
16:52not a really great hook. So, I was
16:54really impressed with the quality of the
16:56ad headlines. Similarly, the email
16:58newsletter prompt, where I asked it to
17:00write an email announcing the release of
17:02this book, was pretty good as well. It
17:04did not overdo things. It kind of
17:07covered all of the main points in
17:09exactly the amount of succinctness that
17:12I would want without going overboard.
17:13And it's pretty good. Like, "I've got
17:14news I've been sitting on for a while
17:16and I can finally share it. My new epic
17:17fantasy novel, The Crown of Hunger,
17:19releases on December 21st. This one has
17:21been a labor of love. Imagine The Count
17:22of Monte Cristo, but dropped into a
17:24brutal fantasy world where betrayal is
17:25only the beginning of the protagonist's
17:27problems. Here's the pitch. And then it
17:29gives us a little bit more in-depth
17:31right here. If you've been waiting for a
17:32revenge story with teeth, this is the
17:34one. Pre-order The Crown of Hunger here.
17:36Thank you as always for reading. I can't
17:38wait to put this book in your hands.
17:40Signed, your name. All right, and that
17:41was pretty good. All right, last but not
17:42least is my SEO article super prompt
17:46where I give it a massive prompt that
17:48has examples of articles about how to
17:51write a fantasy novel. I ask it to write
17:53one. I ask it to write 4,000 words. Now,
17:56writing a good article is something that
17:58most AI models can do these days. So,
18:01the main thing I'm actually looking for
18:03when I do this is does it seem well
18:05organized? And so far, you know, it
18:08looks like it is. We have a good section
18:10on what is a fantasy novel. It talks
18:13about subgenres here. Common tropes in
18:15fantasy novels, it goes through those.
18:18Steps to write a fantasy novel, and then
18:20in the smaller heading it go it starts
18:23with identify your audience, right? You
18:25know, it's the layout's good. The the
18:27amount of visual variety is good. Uh but
18:31the main thing I'm looking for is did it
18:33write roughly the 4,000 words that I
18:35asked it to write in the prompt? And in
18:38this case, it did not. It wrote 3,149,
18:42which is close, but not quite. Now, if
18:44we compare that to Opus 4.6
18:47with what it wrote for me, that one
18:49wrote 4,500. So, maybe this is just the
18:52the prompt in this case it just didn't
18:53do it. But for some reason in this case
18:55it seems to have gone back a little bit.
18:58That said, that could just be because
19:00it's being a little bit more succinct.
19:01Maybe the writing has improved. But
19:03still, it's more than 1,000 words less.
19:05So, that's not what I asked for.
19:07Regardless, it's still a decent article
19:09and I'd be okay publishing something
19:11like this with a little editing, right?
19:13And those are all of my prompts. So, as
19:16far as my own bottom line for this, I do
19:19think it's an improvement on Opus 4.6. I
19:21know a lot of people think that Claude
19:23was throttling 4.6 somehow, making it
19:26worse so that when they released 4.7 it
19:28would feel like this big leap. My
19:30initial prompts that I did the test that
19:32I did for 4.6, I did back when 4.6 came
19:35out, and I've since compared it to 4.7
19:37and I'd say definitely for sure 4.7 is
19:41an improvement over 4.6. Usually, I
19:43don't see a huge improvement. Usually,
19:45it's just like, "Eh, it feels kind of
19:46the same, but sure, like there's nothing
19:48really bad about the new version, might
19:50as well upgrade." I do actually believe
19:52that this is a significant improvement
19:54over 4.6 in most areas. The only area
19:57where I thought it actually got
19:59potentially worse was in the outline and
20:01the the SEO article super prompt, which
20:04keep in mind, those are some of the
20:05biggest prompts that I have, so perhaps
20:08that factors into things. But given
20:10especially how improved
20:13the scene was when I gave it all of that
20:15style info and everything, that to me
20:17was a big one. To see it actually
20:19conform to the prompt a little bit more
20:22was important to me and follow style and
20:24all of that. So, that's huge one for me.
20:27I personally don't put any faith into
20:29these claims that these AI companies are
20:32making their old models worse in order
20:34to make the new ones look better. I
20:36certainly wouldn't put it past some of
20:37these companies to do something like
20:39that, but the reason I don't think
20:40that's actually happening is because I
20:42always go back and look at what I
20:44originally got when I ran these tests,
20:46and indeed, it was as bad as people are
20:49saying it is now back then. But the
20:51thing is, we've had all of this time to
20:54get used to that model, so suddenly it's
20:56not going to feel as special anymore,
20:58and that's why I think it's the shifting
21:00baseline syndrome, where our baseline
21:03for what we think is quality keeps going
21:05up, and therefore, as it goes up things
21:09are going to start feeling worse. So for
21:10that reason I don't actually put that
21:13much faith into it because it doesn't
21:14jive with the experience I have looking
21:16back at past tests and being like, "Man,
21:18why did I think this was good at the
21:20time?" So that's just something to keep
21:22in mind, but I do think this was a
21:24a solid model. And certainly if you've
21:26been using Opus 4.6 in the past at any
21:29step in your writing process, I think
21:31you will not be in danger of upgrading
21:33to 4.7 at least in most cases. Be sure
21:36to do run your own tests and we'll see
21:37how this thing does with long-term
21:39testing, but for now I think this is
21:40great. So if you like this, be sure to
21:43check out my membership below where we
21:45have the entire writing process
21:47available for you to check out and lots
21:49of genre guides to make sure you get
21:50started with the right style and the
21:52right tropes and everything to get you
21:55on your best feet when it comes to
21:58actually publishing your books. So go
21:59check that out and I will see you in the
22:01next video.