Full transcript
0:00All right, folks. Today we're going over
0:01another end-to-end automation that I've
0:03built. This is one I've been incredibly
0:05excited to do. And honestly, this is
0:08probably the automation I've spent the
0:10most time on just to test it and make
0:13sure it's working. I literally spent
0:16hours and hours and over $100 of tokens
0:19just to test this thing out uh because
0:22this one's really important to me. The
0:24automation in question is a line editing
0:28automation that also doubles as a AI-ism
0:31remover or as I like to call it, the
0:33desloppifier. So, let's dive in and take
0:36a look at what this automation looks
0:38like. So, the way this works is you take
0:40in a document that's already pre-written
0:42and we take this and we'll create a new
0:45document where the new and improved line
0:47edited story will go and then we just
0:51run this and it goes on every 1,500
0:53words or so. So, let's just run through
0:56this one step at a time.
0:59We've got the execute button. This is
1:01all you have to do is just click the
1:02button. Of course, before you click the
1:04button, you do need to make sure that
1:07you have identified the correct Google
1:10uh files for this. So, there's two files
1:13here. This one is the blank one. It'll
1:16be blank when you start and it's where
1:18the the improved version of the text
1:21will go. And then this is the one this
1:23will be your original document that you
1:26want to have edited, okay? Uh then what
1:29these two things this is just a little
1:31bit of vibe coding that I did here that
1:33all these do together is it essentially
1:37converts it into markdown cuz it was
1:39downloading HTML. That can be really
1:42like complex and so we want to like
1:44simplify it a little bit here, which is
1:46what those do. And then this little
1:49piece of code splits all of that into
1:511,500 word chunks. Why 1,500? Cuz I
1:55tested this and I tested this and I
1:57tested this. I tried it with 1,000. I
2:00tried it with 2,000. I find that it
2:02works better with smaller chunks. You
2:04can do it with larger chunks, but I felt
2:06like the quality of the edits were not
2:09quite as good. But I wouldn't want to go
2:11too low like 500 words or so because
2:15that would actually increase the cost of
2:17doing this whole process cuz you're
2:18running it more times. Even though it's
2:20on a smaller amount of words, you're
2:21still running it more times, which means
2:23your costs are going to escalate. I
2:25found there was not that much of an
2:27increase in cost going from 1,000 to
2:302,000, but there was a little bit of a
2:32diminishing returns in uh how efficient
2:35the edits were. So, I found 1,500 to be
2:38about the right level here. So, it
2:41chunks your manuscript and this could be
2:44a book of any size into 1,500 word
2:46chunks and then it starts a loop and
2:47runs this loop on all of the different
2:50chunks of text. And basically, I've
2:53divided this into three specific
2:56sections. We have these two, these two
3:00and these two. And what these are doing
3:03is it's running three specific prompts
3:05to try and identify issues and then the
3:07second one is just a revision prompt to
3:11reproduce the text with those revisions
3:14implemented but not changing anything
3:16else, hopefully.
3:17So, that's what these do generally. I
3:20originally tried to make this as a much
3:23longer string where
3:26each of these first sections would go
3:31into more detail on one specific thing
3:34and then I would try to use really
3:36low-cost models to do it hoping that by
3:38keeping the parameters very low, it
3:40would be able to identify them all very
3:41correctly and give good input on what
3:45the alternatives should be for each
3:47phrase that I it identifies.
3:49Unfortunately, I found that that wasn't
3:50really working.
3:52Um I found that the small models just
3:54weren't that smart even at that really
3:57highly contained level. And so, I found
3:59that I needed to use
4:01stronger models but I didn't want to
4:03have like 15 different steps all using
4:05stronger models because that would make
4:07it unusable in terms of how costly this
4:10thing would be to run cuz remember like
4:12if you've got 15 steps and you're
4:13running that on every 1,500 word chunk,
4:16that's going to add up. As it stands,
4:18these
4:19six steps split into three sections
4:23still costs a fair amount
4:26but this was the the best compromise I
4:30found in being able to do this. Now, I
4:33am working on some other alternatives
4:35that might not even be in an automated
4:39form that might actually be more
4:40efficient for this sort of thing. Uh
4:42there are ways you could do this in
4:43Cloud Co-work for instance that might be
4:46a little bit better. We're working on
4:49implementing a feature that could do
4:50something like this into
4:53my
4:54the Story Hacker tool that I'm
4:55developing. But regardless, this is just
4:58the first step and it this does do a
5:00good job. I think it does
5:02improve the text significantly. But
5:05we've got we just got to remember that
5:07editing is one of the harder things for
5:08AI to do. Ironically, it's easier at
5:11generating text cuz this is generative
5:14AI than it is at revising text. But
5:17regardless,
5:19this is the best thing that I've got so
5:21far to do this. So, let's actually take
5:24a look at the different prompts. There's
5:25like I said, three sections. We have
5:28sentence and paragraph pacing. We have
5:30various line edits and the desloppifier.
5:32And then with each step, we have a
5:34rewrite where it kind of implements
5:36those changes, right? So, the sentence
5:38and paragraph and pacing, this catches a
5:41couple of AI-ism type stuff especially
5:44in terms of words and phrases that are
5:46overused. But this is also just a good
5:49way of making sure that your your story
5:52is is not too monotonous. Like you're
5:55kind of mixing up the cadence and things
5:57like that. So, it goes like this. You
5:59are professional line editor
6:01specializing in prose pacing and rhythm
6:03as well as to screen for AI writing
6:04patterns related to low burstiness,
6:07which is this idea of like sentences
6:09being very equal.
6:11Uh your task is to analyze the provided
6:13text for sentence and paragraph
6:14variation issues that affect pacing.
6:17Perform this task from a line editing
6:18perspective where we are modifying
6:20sentences and bits of dialogue but not
6:22changing anything significant about the
6:23meaning or flow of the narrative and not
6:25changing substantial amounts of text.
6:27This is strictly an analysis and
6:29planning task. Do not rewrite any of the
6:31text. Do not suggest new sentences. Your
6:34job is to identify problems and produce
6:35a precise improvement plan that can be
6:37implemented later. Follow this plan. And
6:40then we have a bunch of steps for it to
6:41follow. So, identify areas where shorter
6:44punchier sentences are needed to speed
6:46up the perceived pace. This should suit
6:48panic, shock, fear, revelatory moments,
6:51and sharp emotional beats. Other than
6:53the moments in step one, the rest of the
6:54narrative should be a healthy mix of
6:56short, medium, and long sentences that
6:57rise and fall with the emotional
6:58trajectory of the scene. Identify areas
7:01where this could be improved. And then I
7:03keep going on
7:04about different ways that this can kind
7:06of manifest inside of the text. Uh so,
7:10we have like formulaic paragraph
7:12architecture. We also flag any of the
7:14following words words or phrases that
7:16act as transitional and filler language
7:19or abstract language where concrete
7:21details should be etc. Examples include
7:23and we have a whole list here like uh
7:26you know, furthermore, consequently,
7:28therefore. Uh needless to say, it was
7:30worth noting you know, these are all
7:32kind of superfluous phrases that you
7:34don't really need most of the time. And
7:36then we also have a bunch of other
7:38phrases here that are very AI-ish like
7:42is a testament to key pivotal moment
7:46reflects the broader setting the stage
7:49for, deeply rooted, I see that one a
7:52lot. And I just keep going with this
7:54list until we get a whole bunch of
7:56different things. We got delve, got
7:58bolstered. Boasts, I see that one a lot
8:03and just keep going with that.
8:05And then we have identify any sentence
8:07that makes a generic claim about the
8:08subject
8:11that could equally apply to dozens or
8:13hundreds of similar subjects without
8:14modification. And then we finally wrap
8:17this up with make a plan to improve on
8:19the pacing of the scene to fix all of
8:20the issues above. Create an improvement
8:22plan or change log on how to improve the
8:24text. For each identified issue, quote
8:26the sentences or passages involved,
8:29state the problem
8:30and clearly label the recommended aka
8:32the exact word or grammar or sentence
8:34structure to cut or add or change. Only
8:37mention those issues that need fixing.
8:38And then of course, I always wrap these
8:41prompts up with remember to be extra
8:42thorough to make sure you've got all the
8:44instances of these issues. Output only
8:46the improvement plan and the change log
8:47with the identified problem and the
8:49suggested fix for each issue. Quick
8:51pause. If you're an aspiring author who
8:53wants to write multiple legit books
8:55using AI without sounding like a robot I
8:58run a private group called Story Hacker
8:59AI. Now, it's currently closed, but if
9:02you join the wait list using the link
9:03below, I'll immediately send you my full
9:06prompt pack for free. These are the
9:08exact long-form prompts that I use to
9:10outline, draft, and edit books. Many of
9:12them are hundreds and even a thousand
9:14words long. So, not those one-line junk
9:17prompts that you see from tech bros on
9:19YouTube. To get them, follow the link in
9:21the description, drop your email, and
9:23I'll send the prompt pack straight to
9:24your inbox. You'll also be the first to
9:26know when the group reopens. And now
9:29back to the video. And then this step
9:31goes and does a rewrite based on what
9:33this one gave it. So, that one was kind
9:35of a mix of line editing and uh AI
9:39desloppifying all related to sentence
9:43structure for the most part. This next
9:45one is getting a little bit more
9:47granular on other line editing tasks.
9:50So, it starts out pretty much the same
9:51as before. You're professional line
9:54editor specializing in creative fiction
9:55and narrative nonfiction.
9:57And then I have a follow this plan. So,
9:59we have instruction here for adverbs.
10:02Uh and I've got a bunch of instruction
10:04on what good adverbs look like and what
10:07bad adverbs look like and, you know, how
10:09to identify the ones we want to remove
10:11or change. And oh, yeah, here's here's
10:13the one on good adverbs right here. Then
10:16we have one on dialogue tags.
10:18Uh this one was surprising to me when I
10:20learned it, but I realized that I've
10:21learned that most of your dialogue tags
10:24should be said or asked. Some people
10:26will say, I even had an editor tell me
10:28this once,
10:29uh that you want something a little bit
10:30more exciting than that, but that's
10:31actually not true. Like you can look it
10:34up. Anything that is not said or asked
10:37actually tends to jolt the reader
10:39because it's not what they're expecting.
10:41The reader is expecting to see said or
10:43asked and their eyes just kind of like
10:45flow over it, right? It's just like they
10:47barely
10:48uh register it just to kind of know
10:50who's talking, right? That's all that
10:53the dialogue tags really do is to let
10:55you know who's talking. I will
10:57occasionally use dialogue tags like
10:59whispered, but for the most part you
11:00just use said or asked.
11:02So, this one checks for that. Then we
11:04have action beats, which is another
11:06alternate way of mixing in some of the
11:10action of the scene with the dialogue in
11:12a way that means that you don't
11:14necessarily have to have dialogue tags,
11:16which can also be good.
11:18Instances of under or over tagging.
11:20Sometimes you don't need as many
11:21dialogue tags cuz it's very clear from
11:23the flow of the conversation who is
11:25talking.
11:26Uh but sometimes it's not, so
11:29we have a check for that. Passive
11:31indicators, so this is looking for
11:33things like passive voice and I have a
11:34bunch of different examples of what that
11:36looks like. Here, I'll scroll through
11:38them.
11:39Let's see. And then we have cliches and
11:41I have a whole bunch of cliches that
11:43I've listed including but not limited to
11:46stock emotional idioms, phrases like
11:48heart pounded in her chest or couldn't
11:49believe his eyes, blood ran cold,
11:51stomach dropped, butterflies in her
11:53stomach, lump in her throat.
11:55You know, all of these cliches that have
11:57built up over time. Those are all
11:59mentioned here and I have it look for
12:00those. And then of course redundancies,
12:02identify all redundancies including but
12:04not limited to
12:05internally redundant word pairings,
12:08phrases where one word already contains
12:09the meaning of the other such as added
12:11bonus, end result, close proximity,
12:15uh past history, etc., basic
12:17fundamentals.
12:18Uh and then a bunch of other things to
12:20look for there. And then repetitions,
12:21identify all word repetitions including
12:23but not limited to filler and junk
12:25words.
12:26Uh things like just, very, really,
12:28quite, still, so, suddenly, that, etc.
12:31Uh and just going through all of that
12:34and including um
12:36words that are repeated frequently by
12:38the author that are maybe overused, uh
12:40which is a common issue with a lot of
12:42authors, especially beginning authors,
12:45who will often rely on certain words
12:47over
12:48uh like more than they need to.
12:50And then we wrap it up with the same
12:51kind of ending here, make a plan to
12:53improve the prose, etc., etc.
12:56And then in this one it all gets
12:58rewritten.
13:00All right, and that leads us to the last
13:02one here, which is the desloppifier.
13:04And this one's really cool. Um this
13:06might actually be really educational for
13:08you um just to know what to look out for
13:10when you're going through and uh editing
13:13text. You know, what kind of things AI
13:15does a lot. We already covered a bunch
13:17of those AI words in the first step
13:19here,
13:20but there are also specific ways of
13:23talking that it uses a lot. All right,
13:25so this one starts out similar to the
13:26other ones. You are a skilled line
13:28editor tasked with identifying and
13:30flagging AI generated writing patterns
13:32in a piece of text.
13:34I use this one last just to make sure
13:38uh there are no steps ahead of it that
13:40could reintroduce some of these issues.
13:43Um so, follow this plan. Negative
13:45parallelisms. This is probably one of
13:48the most common AIisms that you'll see a
13:50lot. Scan the text for parallel
13:52constructions that set up a contrast or
13:54correction as though the author is
13:56preemptively challenging an assumption
13:58that the reader might hold. Flag every
14:00instance of the following patterns and
14:01quote the sentence or passage.
14:03Not just blank, but also blank.
14:06Or it is not just about blank, it's
14:08blank. Or not X, but Y.
14:11Or not X, not Y, just Z. So, stuff like
14:14that. We also have the rule of three.
14:16Scan the text for the overuse of triadic
14:19structures. Flag every instance in which
14:21the author lists exactly three items,
14:22whether adjectives, nouns, or short
14:24phrases, in a way that feels formulaic
14:27or padding rather than necessary.
14:29Um so, examples include adjective,
14:31adjective, and adjective or noun, noun,
14:34and noun. Short phrase, short phrase,
14:36and short phrase. So, it'll look through
14:37and find those. Uh of course, here's a
14:39big one, overuse of em dashes. AI
14:41generated text overuses em dashes. While
14:43em dashes have legitimate uses, AI
14:45writing inserts them in places where a
14:47comma, parenthesis, colon, or period
14:50would be more natural and appropriate.
14:51Identify every em dash in the text and
14:53flag instances where
14:55a comma would be a more natural choice.
14:58Parentheses would be more appropriate
14:59for an aside or clarifying phrase. A
15:01colon would better introduce a list or
15:03explanation. Two em dashes are used to
15:05set up a phrase that would read more
15:07naturally in parentheses or between two
15:08commas,
15:10etc., etc. And I actually added a note
15:12here. I just said, "Always remove em
15:14dashes, all of them."
15:16Uh not because there aren't reasons
15:18where an em dash would work, but in this
15:21case I find that even if I tell it to
15:24remove all em dashes, some will still
15:26get through. And so, I want to be okay
15:28with a few getting through rather
15:30because if I just say, "Try to find
15:32appropriate places for em dashes," it'll
15:34still overdo the em dashes. So, by
15:36saying don't have any em dashes, I'll
15:37just get a few and that's all I need.
15:41Uh all right, then we have collaborative
15:42or correspondent language.
15:44Um
15:45the this is basically looking for uh
15:48moments where the text says like, "Oh, I
15:50would be happy to do that for you." and
15:52stuff and we want to strip any of those
15:54out.
15:55Uh safe and predictable word choices, uh
15:58high priority watch list. Uh so, here we
16:00are again going through a couple of
16:02very common AI words. Then we have
16:05abstract language where concrete details
16:07could exist. Flag any sentence or phrase
16:09that describes a feeling, quality,
16:10atmosphere, or concept in general terms
16:12without grounding it in something
16:13physical, observable, specific, or
16:15sensory.
16:16So, we have like examples of things to
16:19flag, emotional generalizations like she
16:21felt overwhelmed, he was overcome with
16:23grief, uh etc.
16:26Then we have safe flat transitions. So,
16:28this things like moreover, additionally,
16:31furthermore, in addition, in addition to
16:32this, not only that, as a result, you
16:34know, these kind of like wordy
16:36transitions that are not necessary.
16:38Next, we have unearned personifications
16:39and metaphor. Flag instances where an
16:41object, abstraction, or non-human thing
16:43is given human emotional qualities in a
16:45way that feels decorative or automatic
16:47rather than earned. We see this a lot
16:49where it gives things that are kind of
16:50inanimate some kind of anthropomorphized
16:54uh quality to them or metaphor.
16:57Um so, things like um the silence spoke,
17:01right? Or even common phrases like a
17:03weight lifted or the walls came down.
17:05Then we have a make a plan to improve
17:07everything, create the improvement plan,
17:08etc. And then of course, it runs through
17:11and rewrites the thing here.
17:13Uh then what it does is it's gone
17:14through and it's done that whole thing
17:16and has rewritten it three times. I
17:18should point out that my rewrite prompt
17:20is specifically designed to
17:22I don't use the term rewrite. I say,
17:24"Use the text of the original text in
17:26the improvement plan. I want you to
17:27implement the suggestions in the
17:29improvement plan.
17:30Only implement the suggested changes and
17:32do not change anything else about the
17:33original text. Reproduce the entire text
17:36with the suggested changes made." So,
17:38what I'm doing is I'm only asking it to
17:40make those very specific changes and uh
17:42in all of my testing, thankfully, it
17:45does usually maintain the overall flow
17:48of the text in the paragraphs just
17:49changing those very small
17:51uh improvements that it can make. So,
17:53then it adds all of that to the document
17:55and then repeats the loop again on the
17:57next 1,500 words and keeps going until
17:59it's done. So, um kind of a lot going on
18:02here. I do think there are more simple
18:05uh effective ways to do this. It's it's
18:06something that's kind of top of my
18:07priority list right now, so I would love
18:09to hear in the comments uh the ways that
18:12you have done it. Have you worked with
18:14like a skill in Claude Co-work? Uh do
18:17you have uh some something you vibe
18:19coded that could do something like this?
18:21I really want to get this one right
18:23because this is one of those things I
18:25think is really, really needed in the AI
18:27writing community and in the larger
18:29writing community as well. So, let me
18:31know down below and I will see you in
18:33the next video.