Transcripts
1. Introduction: Have ever felt that your
thoughts are raising ahead of your fingers when you're trying to
write something, you are definitely not alone. But these days, we're actually living in a time where tools that used to seem
like something out of science fiction are real. Now, you can sit down in
front of a blank screen, start talking, and watch as
your words appear instantly. I'm Robert, and I'm
all about finding practical workflows
that save time and keep your attention on
the work that matters. In this class, you
will learn how to use AA dictation to do more than
just simple voice to text. Spent serious time diving
deep into this space, even writing most of the official documentation
for Superwhisper, my top Mac app for
this kind of work. We will combine fast, accurate transcription with AI, making sure your tastes and
style stay completely yours. So, what is this for? Well, if you are
someone who deals with text related tasks
on a daily basis, whether you need
to answer emails, write articles,
online video scripts, enjoy creative writing or
even for those who are into note taking and methods for
quick capture of knowledge. I have structured this class
in four levels of dictation. We will start with the basics, understanding how
these tools enable you to get clean,
accurate transcriptions. We will quickly move into using
AI to clean up your text. Then we will refine
your writing style. You will learn how
to prompt AI to improve clarity,
tighten sentences, and make your content sound more natural or while keeping
your authentic voice. Finally, we will build
specialized workflows that turn dictation plus AI into
a real writing partner, one that follows instructions, adapts to your context
and helps you move fast. You don't need to
be an AI expert. As for tools, my
main recommendation is Superwhisper or Mac. I will also show you
how to set things up with spokene and voicing. You can use any AI dictation
app you like as long as it lets you customize the prompts that
process your text. Ideally, your chosen
app can also read context from your current window for more advanced workflows. Core ideas we will
cover will work with any tool that
meets those basics. By the end, your grading
process should feel faster, smoother and more enjoyable, more thinking and creating, less grizzling with mechanics. If you're ready for a faster way to get words out, stick with me. I'm excited to show
you what's possible.
2. Overview & Basic Concepts: I want to take a minute
upfront just to lay out more details about this class and give you the big picture. We will walk through
what's coming up, talk about some key concepts
we're going to cover, and touch on things that will help you get the most out
of our time together. Yes, AI dictation is an
incredible productivity booster, but there truly are so many ways in which we
could cover this topic. And because of
that, I think it's important that we start
on the same page. Not only do I want to teach
you how to use these tools, but I also want to share my personal perspective on some of my creative framework that comes into play anytime I use AI. That way you have got a sense of where we're heading
before we go further. Now, before we jump in, it's important to
clarify that this is not about replacing typing
or making it outdated. Typing for many people,
including myself, is still a huge part of
the creative process. Under some specific use cases, there's something about
putting your hands on a keyboard and
letting your thoughts develop one word at a time
that can help you see ideas, take shape, and shift as you go. Same as handwriting,
there's a kind of thinking that can only happen with
these more physical methods. Sometimes the slower
pace actually lets you find unexpected connections and refine your thinking
in real time. So no, manually typing
isn't going anywhere. But talking your ideas brings a different
set of benefits. There's a kind of freedom
you get when you speak, especially when you're trying
to capture ideas quickly, brainstorm or just get
past the mental block. It's also extremely useful for quick communication
skill that I'm sure we have all had to develop
in some way or another. What's new now is that EI and
modern speech recognition actually let us use this way of working in a way that's
extremely practical. We didn't have that before. AI dictation gives you another
way to get your ideas out. And once you learn how
to make use of this, you can switch
between talking and typing depending on what
the situation calls for. You can truly achieve a level of flexibility that was
not possible before. And if you understand
this well enough, you don't even need to sacrifice giving up control
or losing your voice. Here's what I want
to dig a bit deeper. It's not just about getting comfortable with
dictation as a tool. It's about having some
understanding what's actually happening behind the scenes as a creative and specifically, when dealing with AI, trying
to get a feel for how all of this works is that
you can actually steer these tools so they
don't end up steering you. Because honestly, it's not just your words being turned
into text anymore. Now in this new generation of apps and technology
we have available, everything you dictate
can also be filtered, formatted and shaped by AI. These massive language
models working in the background that decide what your words will
look like on the page. Something I have noticed
is that a lot of people feel
intimidated by all of this AI stuff or
they worry that it's going to take over art
and human expression. I think these are
valid concerns, but I like to look at this from two angles, creativity
and productivity. These tools can save
you time, real time, which means more energy to the creative work that
actually matters to you. More you understand them, the less you have to fear
and the more you can focus on what you want to say and
how you want to say it. That's why I care
so much about this. It's not just about
learning one specific tool, but to really get how
the whole process works, how prompting affects what the AI does and how you
can control the output. The more you understand the
logic behind these systems, the more you can pick
up any new tool, tick it to your taste,
and feel right at home. I don't want you to feel
like you have handed over all your creative
decisions to an algorithm. You should always be able to
bring your own perspective, your own style, and your
own words to the table. We go through the
lessons, we will move from simple
to more advanced. I will walk you through my own process for
creating proms. Real proms I actually use plus some principles that you can follow to adapt them
to your own needs. But the goal isn't just for you to simply copy and
paste what I give you. I want you to start
thinking about how a adictation can fit
your own specific needs. So every time we cover a new
level or type of format, I point out why I prompt the way I do and what
I am keeping in mind. That way you are getting
both the how and the why, not just the words to type in. Now, if you would
like to make this even more practical,
here's an idea. I want you to come up
with something personal, something that matters to you
and throughout the class, design a prompt
you'll be using for that specific use case
when dictating with AI. It can be a prompt that
you adapted from one of the lessons or even something that AI helped you come up with. We will learn some useful
things about prompting, but the main thing is that it's actually connected to something you want to solve or improve through you
learn in the class. By the end, you'll
be speaking and getting better results
with less effort. Very happy you are here.
Let's get started.
3. Voice to Text - The Foundation: It used to be that dictation felt like more hassle and help. You would use the built in
options on your device, but they would miss
half your words or you would end up
with a mess that took longer to fix than if you had just typed
everything out yourself. Specialized dictation software
was sometimes better, but it came with its
own learning curve. You had to spell
out every comma, every period and
say new paragraph just to get any
kind of structure. It kept a lot of people
from using voice at all, but now things
have changed fast. There are two big
reasons for that shift. The first one is the rise of new AI models for
transcribing speech. These are way more accurate and a lot faster than
what we had before. One open source
model by the name of Whisper changed
everything making this technology
available for free or at a very accessible
price for everyone. Currently, there are a lot of other options that have
different advantages, and I'll talk about
that in a moment, but I want you to
know that WISPR continues to be a very
solid alternative. The second big shift
is the introduction of large language models or LLM. Are the same kind of AI models powering tools like ChatGPT. There's many different
tools that started to use LLM to clean up format
or rewrite texts. We even have the
native implementations like Apple Intelligence. But something that people
have started to notice are the possibilities that
appear when combining both. AI transcription models
together with LLMs. You just speak naturally and let the AI
handle all the rest. But let's slow
down for a second. Before we go into how LLMs
can shape your grading, it's important to understand the basic job of a transcription model
and its possibilities. This is the AI that listens to your voice and transforms
it into text in the first. That's it. It's not trying
to understand what you mean. It's not trying to organize your ideas and it's not
fixing your grammar. It's just turning
sound into words. Once you download one of these dictation tools
that I recommend, you may find that there's a
lot of models to choose from to the point that sometimes it can get quite overwhelming. Some models can work offline
right on your computer, which is great if you
care about privacy or don't want to send
your audio to Cloud. Others work entirely online. No downloads. You don't
have to worry about your computer space or if
your system can handle it. If you're looking
for local options, you will see that there's
some models that are smaller in size, often
called distilled. This may be very light and fast, but usually you
give up something like multilanguage
support or accuracy. Bigger models are usually better with things
like punctuation, handling multiple languages,
detecting accents, being able to
understand when you are whispering or are in
a noisy environment, but they can be slower
or need more resources. That is why cloud transcription is still a good option since you are letting a server
take the load and simply provide
you with the results. Every app manages transcription
models differently. My GT app currently
is Superwhisper. You just have to
dive in the models, and little badges tell
you whether a model runs in the cloud or
needs to be downloaded. In the same place, it
tells you if the model supports multiple
languages or only English. If your chosen model actually supports
multiple languages, you can often select that
model's language specifically, or you can let the model to detect the language
that you are speaking. With Spokene which would
be my second app of choice for getting into everything that we'll be
learning in this class. You have a specific tab to
manage your dictation models. Difference here is that if you are using the
application for free, you have to get
your own API keys from the different services that are included with the app. If you decide to pay for it, you don't have to
use your API keys. Here, we also have
information about speed, some metrics about accuracy, and the languages supported. When you use the transcription
model by itself, you're only going to get the
raw version of what you say. Some models can handle
punctuation pretty well, but they will still
not understand the structure of
what you're saying. So the question here is, when should you use transcription model without
passing it through an LL? I think there are
lots of moments when speed matters
more than polish. For example, if
you are dictating notes to yourself,
brainstorming, writing a rough draft or sending a quick
message to a friend, you probably don't care about the perfect grammar
or formatting. In those cases, simple
transcription is perfect. The quality of the
results that you get from different models is still important to
consider, of course. Let me start dictating. Right now, at the time of
me recording this class, one of the fastest options
that we have available is a local model called
parakeet V three. It is almost instant. It supports multiple languages, and it has the advantage that it's local and it has
a very small size. You can see that I
would stop talking and immediately my text
is already there. The other hand, it
has several downsides and one of them is the accuracy. I'll just make mistakes when your pronunciation
is not very good and the punctuation
may sometimes be way off. Let me switch to a cloud model. Now I will dictate again. In Superwhisper, we have
this option called ultra. I know for a fact that
this is a whisper model, which means that it comes from that first generation of transcription models that revolutionized the
entire process. In this case, it's a distilled version with
multiple language support. It's slower than the
one I was showing you, but the accuracy in both punctuation and word
recognition is much better. It's not perfect,
but if you care more about quality than
speed, it's good enough. The downside is that
is a cloud model, so it will depend in whether you have
Internet access or not. Now, if you care about privacy and want to keep
everything local, you also have other options. Overall, I would say that
ultra turbo V three is a good one better than the two previous that
I just showed you. But again, it will depend a lot on your system
specification. Holy, the accuracy of many of
these transcription models, even without an LLM pass
is getting pretty good, especially if you take a little extra time
to set things up. Here's what I want to mention about something
called prompting. You may have heard how prompting works with large
language models, and we will also dive into
this in some other lessons, but with transcription models, and specifically
with those based on whisper, it's
something different. Like I have told these models just listen and turn
sound into text. They don't really understand
what you're saying. But you can still give them a little help to make
them more accurate. Think of a prompt here as a
group of hard to spell words, names, acronyms, or technical
terms that you might use. Prompting for transcription
models may also help with guiding
language detection and even punctuation. There are still
many other factors that come into play here, like if it's a distill
model or not or if it's a model based on whisper or
something else entirely. So models don't
even have this like the crazy fast packet model that I showed you a moment ago, different applications
may refer to the transcription prompt
in a different way. In voice ink, for example, we have an advanced setting
that calls it output format. This is applied whenever you choose a whisper
model and what you should write here is something like my dictation may
include the following names. Robert Albert William. What am I providing with
this simple sentence? I am indicating my
language, which is English. I am indicating that
I need punctuation because I am
including the column, the commas, and the period. And I am also
including some names that I need written
in a specific way. As this technology
continues to advance in an attempt to make
setup easier for users, many apps provide a more
automated way to do this. Since I am here on voice ink, I want to show you this tab
that is called dictionary. Here you can set
word replacements with correct spellings. If you start to dictate and
went up and after a while, you notice a pattern
in which your words start to be incorrectly
spelled, you can fix that here. For example, if I notice that Robert is always being
spelled like Robert. The correct spelling section is a list of words that will be passed later when the
transcription is processed by AI. In Superwhisper,
everything is handled in one tab called vocabulary. My intention with explaining
all of this is also to help you troubleshoot because in the case where you
have, for example, selected automatic
language detection for transcription and you
are speaking in Spanish, but your result is being
returned in English. Well, one of the things
that may be affecting this is a prompt that is
being passed to the model. With Superwhisper, one
successful way in which I add some counterbalance
for language detection is inserting a few
words in Spanish. Bottom line is this, depending
on the app you are using, you might see
different ways to help these transcription
models do a better job. But the idea is the same. You are not getting the
deep creative steering you have with LLMs, but you can absolutely improve some of the
quality of the results. It's one of those
small twiks that can save you from a lot
of annoying mistakes, and perhaps it can also
save you a lot of time by not having to use AI processing for
simple dictation needs. Personally, at this point in
time for quick dictation, I am mostly using the parakeet
model with Superwhisper. When I feel I am not getting the accuracy I need
for a specific task, I switch to a mode that adds
a very quick cleanup on top. This may change at
any time because this technology is
improving very fast. We will learn more about
this in the next lesson. For now, what I want
you to do is that you spend some time trying things
out in your app of choice. Think about whether
you need a model that keeps everything
private and local or if you are fine
using a cloud based option, decide what matters more
for you, speed or accuracy. If you're working with
languages other than English or often use unique names
or technical words, then you may need to pick an app that lets you
customize vocabulary. You may need to go for something
that is not super fast, but that will work
best for your case. This is also a good
time to set up your vocabulary or replacements
if your app supports it. Run some quick test and see what feels most
comfortable for just dive in and
play around with the options until you find
something you are happy using. That's the best
way to get a feel for how all of this
actually works. So that's a foundation, getting your voice
on the page and understanding some
of the choices that will be presented to you. In the next lesson, we
will look at how LLMs can take that row text and turn
it into something sharper, more readable, and
ready to share. Stick with me and let's keep building on what
you have learned.
4. General Cleanup - Making it Readable: Alright, welcome back.
So in the last lesson, we talked about that very first
level of voice dictation. The one that gives you a raw, simple transcription is
usually the fastest. And honestly, a lot of these
AI transcription models do a fantastic job with accuracy and some
basic punctuation. This should be enough
to cover some of the most basic and
quick use cases. But if you remember, I mentioned that with raw transcription, the model doesn't understand
what we're dictating. So you cannot auto insert paragraph breaks or detect and fix whenever
you use filler words. Or if you accidentally corrected
yourself mid sentence, raw dictation still captures
every single sound. This lesson, we
are going to take a really important
step up from that. We will start working
with AI prompting. In other words, we will be taking our raw
transcription and having it processed by a large language model
to give it some polish. This is about making your
dictated words instantly more readable without having to go in and manually fix small
details yourself. And this is a pretty
powerful step because this is a point where a lot of people start seeing the real benefits of
using dictation with AI. Now, a lot of these
AI dictation apps in the market also automate
this step behind the scenes. They will just give you a more polished version of your dictation
right off the bat. Or at least they
will already include one preset or one mode that
would be perfect for this. In Superwhisper, when
you create a new mode, you can simply select message, and it will do this kind of cleanup that I'm
telling you about. In voice ink the moment that
you activate AI enhancement, this is the default
cleanup that happens. Is the same as with message that I showed you
in Superwhisper. If you go with
Spokene time being, there's no presets here and
the process is more menu. You could insert something like fixed grammar
and punctuation. But the idea with this specific lesson that you
are watching is to help you get better results with applications where you face
this option to customize. Because of that, it
is important that I share with you some basic
concepts in prompting. The last lesson I told you about prompting the
transcription model. Now that we want to
get into prompting the large language model that will actually
understand our content, there's one more term that is important to know
the system prompt. The system prompt
usually defines the AIs identity for
the whole conversation. If you tell it, you
are a translator, you will keep acting
like a translator as the messages go
back and forth. Now, dictation apps are
not chat applications. Most of the time you are
doing a single pass. The app will send
your transcript text, the LLM does some
processing on it, you text back and that run is. There's no conversation here. So even though there's still a system prompting
behind the scenes, it applies to that
one interaction. Then the next recording
starts fresh. Now, an important thing to know is that in most dictation apps, the system prompt is
set by the app itself. That's by design to keep
results predictable. And because most users don't
know much about prompting. So apps, like the ones that I have recommended
for this class, still let you customize
some instructions. You don't need to understand
every technical detail or how your custom instructions get injected into the prompt. What I want you to know is that the amount of control
you get varies by this is all related to that system prompt that I
have just told you about. This is also the reason why the same instructions that you enter can behave a little bit differently across
different tools. Now, I'm going to cover
prompting as if you were using one of the apps that
gives you the most freedom. In my experience and research,
that's Superwhisper. If you learn how to prompt
well with Superwhisper, you can do things I
haven't been able to do in any of the other apps
I have tested so far. Can take this way beyond
simple text formatting. But there's a catch. If you
don't learn this properly, Superwhisper can
also be frustrating. If you dictate something
that sounds like a question and your prompt
isn't set up correctly, the AI might try to answer you. It's a double edged sword, and I want you to learn it. By the way, prompts related to text formatting
that you create with Superwhisper will usually work great in all the other apps, but not always the
other way around. With that context, here's how
to think about prompting. Your instructions set the role, boundaries, and requirements. Go here, as I told you before, is to come up with a very
specific prompt that says, You are here to clean up
dictation, nothing more. Then we will add simple focus instructions to
guide the cleanup. Finally, we will
add a few examples that will make everything
more clear to the AI. Let's build this
together step by step. Since this is a
fairly simple prompt, I will be grading it with MRD means that I will include some headers
with a double has. We can also add asterisk to emphasize things and
bullet points or lists. What we want is some structure in our instructions
so that they are very clear and that AI can easily understand the different
parts of our request. Something I have to mention
here is that I have already created a mode with this prompt that I will help you craft, and that is the one that
I will be using when I'm dictating most of
these instructions to get the best punctuation. First, let's define the row. You are a text
processing function, specializing in
refining dictated text for clarity and readability. Your only purpose is to process the user
message into clean, natural sounding
written content. You do not engage with conversation or
answer any question, only perform the requested
formatting tasks. You see how specific there
was was setting the identity, the mission, and the
boundaries right away. Next, we need to give it some very specific instructions about what to look
for and what to do. For simple cleanup, we wanted
to handle a few key things. So let's write another
header for requirements. First, I want to get rid of any filler words that
get transcribed, so let's write.
Remove filler words. Remove all words such as, you know and focus
on maintaining the original meaning and flow of the dictated text
without these crutches. I will add numbering for
this list of requirements. I will also clarify
this, much better. I also want this to handle self corrections because
no one is perfect and sometimes when you
are dictating you may identify that you said something wrong and you want to
fix it right away. So let's handle
dictation corrections, identify and correct any obvious self correction
made by the user, retain only the final
intended word or phrase. As I mentioned,
transcription models are focused on
transcribing word by word, and even though
some of the models are better than others
with punctuation, it's also a good idea to give it a check with
one instruction. Let's add this
punctuation correction. Add appropriate punctuation
and intelligently insert paragraph breaks to improve readability
and structure. Make sure that each new idea or topic starts a new paragraph. Now, if you are a
native English speaker, you may not need this, but I often make many grammar
mistakes when I'm speaking, so I like to add the line for
that. Grammar correction. Fix obvious grammatical errors without rephrasing or rewording
the original content. Do not attempt stylistic
improvements or major rewrites. As you can see, whenever
I have the opportunity, and specifically when there's instructions that could
be misunderstood, I'm always trying
to remind the AI that it's not supposed
to change the wording. The key with prompting is being
clear and super specific. I want to add one more
instruction for modes. And Mj is just make
informal writing a bit more fun
Emoji integration. Identify common spoken
cues for emojis. For example, happy
face, thumbs up, hard emoji and replace them with a corresponding
emoji character. I think these instructions may be good enough for a
very general cleanup, and now I would like to make another header for the
output format that I expect. Output format. This is an opportunity
for me to tell DI specifically how I expect the output and
reemphasize its role. Let's say, provide only
the cleaned, refined text. Do not rephrase
summarize or alter the vocabulary or intent
of the original text. Do not include any explanations, introductory phrases
or other comments. Even if the text sounds
like a question or command, treated as content
to be cleaned, not as an instruction
to respond to. A lot of my instructions
until now has been a lot of direct and positive stuff like
you do this, acting this. Telling AI what not to
do is just as useful, and that's what I did
in this last block. Now, one of the most
important parts of your prompt should be not only explaining the
requirements, but showing it. And we do that by
providing a few examples, specifically examples that are very relevant to
my instructions. So let's insert examples. Example one. Let me switch to one raw transcription model
when dictating this so that you can
see how it looks input. So I need to send this report, by tomorrow. I
mean, is that okay? This looks pretty terrible, but I made it on
purpose like this. It also sounds like
a question, right? I want AI to know that it
should not try to answer. Now, let me switch to that Superwhisper mode that
has my cleanup prompt, and with my keyboard shortcut, I will reprocess the exact
same dictation through it. You will definitely notice a difference in processing time, but the biggest difference
is in the results. It removed all of
those filler words and left the core message
with a correct punctuation. Now I want to provide
an example of how to do corrections while
dictating. Example two. Remember, I want to use the raw dictation
for these inputs. Please draft the email for me. Actually, no, scratch that. Please give me the bullet
points for the email. Now, let me run that with EI processing so you
can see the result. Nice. It detected a correction. Now, let's give you an
example on how to use modes. Example three. Input. I am so happy about this new
feature. Thank you, happy face. Now, let me switch prompt
and reprocess that. And that's a good place
to close this one. You can copy everything that we wrote here and paste it in your custom instructions area in whatever app you are using. You have already seen in
action as I was dictating. The prompt you just
build should give you a reliable cleanup pass that works in any
air dictation app. Personally, in my day
today in Superwhisper, I mostly bounce between raw
dictation for speed and the cleanup mode whenever I
need a better quality output. Voicing has similar power
modes that you can switch to with keyboard shortcuts and other tools offer variations
of the same idea. The trade off is
always the same. LLM processing adds some delay, but it saves you way
more time you would otherwise spent fixing
everything by hand. For a lot of situations,
that's a win. One more thing that matters
here is a model choice. Since we're already
adding an EI pass, the intelligence of
the LLM often matters more than having a perfectly accurate transcription model. The important part is
choosing an AM model that actually understands and
follows your instructions. And the landscape is
truly changing so fast. Right now, GPT five mini is good enough for
this kind of cleanup. Recently, I'm using
Kimi k21 model provided by Grock
which is super fast. Your setup might be different, and that's the point that
I'm trying to make here. You can already start tuning
your dictation up for your system and your use cases.
So here's your homework. Try this prompting in the
app you already using. Test a couple of
different AI models. Try using super
fast transcription, even if it's not
the most accurate, just to see if the AI pass
is good enough for what you. Find the balance
that works for you. And by the way, feel free
to tweak some lines of the prompt to personalize
it more for your use case. If you start testing
around and you find that at some point you get a result that you
are not expecting, something you can also
do is come back to the prompt and add that
mistake as an example, just so that AI learns what you do expect when encountering
a similar situation. In the next lesson, we
will take another step. Instead of just learning
what you just said, we will start asking
DI to improve it, tighten sentences,
clarify ideas, and do more re writes
without losing your voice. It's a bit more creative,
a bit more structured, and I think you
will like how much it improves up your
results. See other.
5. Improvements & Rewriting: Alright, let's make your spoken
words sound even better. In this lesson, we
will talk about how to take your dedication
and prompting AI so that it turns it into more polished, natural
sounding content. For this, we need to have
AI understand what we're trying to say and then help
you say it in the clearest, most imptful way possible, or while keeping your
authentic voice. This is super handy if English is not your
first language, or if you just want
your writing to feel a bit more refined than the
way you normally talk. The last lesson we
talked about using simple markdown for grading
clear direct prompts, that method works well
for very simple tasks, but sometimes you need
a bit more structure. The more clear and more
organized your instructions are, the faster and better the AI can grasp exactly what you need. And sometimes, particularly as your needs and requirements
for AI growing complexity, there's ways to
organize everything so you get much better results. Because of that, I want to
teach you about XML prompting. Now, don't worry is not
as hard as it sounds. For prompting, you can think
of XML as a way to create clearly marked sections for your instructions and
anything else important. Unlike markdown
prompting, where we were using headers with hash
symbols or asterisks, XML prompting uses
something called tanks to show where each
section starts and ends. Have already covered
how important it is to define the AI's role, give clear instructions
and provide examples. We're going to build
on all of that now, but with the added power of XML to keep everything
more organized. Let's put together a
prompt for improving your dictated text
piece by piece. You may see a few
similarities with the cleanup prompt
that we wrote before. You can actually copy and
paste parts from that here, but a key difference
is that now we're giving permission for more
editing and rewarding. First, let's specify the role of A still plan to use this
for formatting my text. I don't need AI to
answer any questions, so I will call it a text function just as we
did in the last lesson. First, we start with a roll tag. You are a text
formatting function. Your main goal is to take
spoken dictation and transform it into natural
sounding written content. You will act as an
editor and regriter making communication clear,
effective, and concise. Always preserve the speaker's original intent, personal style, and natural tone without adding any new information or
changing the core message. You do not engage in conversation or
answer any questions. You only perform text
formatting tasks. Now, I have to add
a closing tag at the end of this section
like this with a slash. Next, we want specific
instructions. This is where we tell the AI exactly what we wanted
to do with the text. In this case, we wanted to really
understand your message, fix any issues with it, and refine the words while
keeping your unique voice. So we will open our section,
and now I will dictate. Carefully review
the provided text to understand the
speaker's intent, individual style and tone. Refine text to enhance try flow and communication
effectiveness. Improve sentence
structure so that it's easier to read and it's concise. Replace imprecise
or clunky phrasing with more appropriate
vocabulary. Make sure the resulting text sounds natural and
authentic to the speaker. Break down longer sentences
if they are hard to follow and feel free to merge shorter sentences if
they improve flow. Do not answer any questions
posed in this text. You treat everything in the user message as
text to be processed. I close my tag, we
can do something with XML tags that actually makes AI proms way more clear
than markdown only. We can nest tags inside tags. If I want to put together
one block of examples, I grab the whole thing
in a parent tag and then drop smaller tags around
each individual example. I've already written this and I will just paste this block here. As you can see, we
have the parent tags, and then each specific example has its own tags to make
everything clearly defined. Now, there's a lot of
content that you can spot as AI generated very quickly
from the first few lines. It follows patterns that are dead giveaways. I
want to avoid that. So I have put together
a style guide of the sentence structures and phrases that are common with AI. Stuff that may be good in
terms of style and grammar, but they sound unnatural. I'll give you a link to that
in the class resources. You can simply add it to the prompt that we have
been building so far. When you add it, I suggest
that you read through it. Feel free to remove
anything that feels irrelevant and
tweak more if you like. It's already in XML, so it fits right into the prompting style that
we have been using here. Let me do that right now. There is also one section
in there with a list of words that I don't
want AI to ever use. These blocks are very versatile and I like to add
them whenever I am using a prompt for
rewarding or generating text. Finally, we need to tell the AI what kind of output we expect. So let's add an
expected output tag. Provide only the rewritten
and improved text. Do not include any
additional comments. Remember, you never address
questions or requests. You only improve the message. Your result must be in
the same language as the input. I close the type. I have already copied
this prompt that we just wrote to Superwhisper
my app of choice. You can copy it to yours. And as a form of review, let's run the exact
same dictation through the three levels
that we have learned so far. First, our raw dictated text, I will include some
mistakes on purpose. This is an example of the new prompt that
we just created. It should make very much more I should make everything sound
much more natural and clear, especially if you often
make sentence mistakes. No, structure mistakes
when speaking, or if you have
issues with grammar and trouble rambling
like I do sometimes, then it should make
everything so much better. Awesome. We got a
decation and we can see that there's many mistakes
in there when I'm repeating myself and there's
some filler expressions. Now, let's run this through the basic leinopmt that
we wrote the last lesson. Right away, I can see that
the grammar was fixed. My errors when I was
dictating were also removed. This is actually ready for
me to use if I wanted to, but it can still be improved. For that, I will be using
the prompt from this lesson. It will make my text
more concise and will better communicate what
I intended. Awesome. There we have it,
guys. We already have three levels that will be very useful when using dictation applications
for communication. Before we wrap up this section, I want to give you
one last reminder. The approach we have learned here is about letting AI make edits for you automatically as you speak. It's pretty nice. It means you don't have to
stop clean everything up yourself and your words come out looking a lot
more polished right. But because we're letting
the AI move beyond just surface level formatting and actually reward
our dictation, there's two important things. First, you may need to use one AI model that is
a little bit more smart and a little bit slower
just to get better results. And second, I still could not
rely on this output 100%. Even when I use something like
this myself or whenever I use AI to help me generate
any kind of text content, I always take a moment to
review what I get back. Will adjust sentences or tweak wording if it doesn't quite sound like
what I would say. It's still my message and
my voice on the line. And I want to make sure
that I'm not adding to all that generic AI content that appears everywhere
online these days. So my suggestion
is that it's best to treat what the AI
gives you as a draft that gets you most of the
way there but plan to read through and make a couple of quick edits if necessary. Hope this lesson on XML
prompting gives you some ideas for expanding what you can do
with AI dictation. To put all this into practice, I suggest you take
your prompt template, use it in your app of choice, and evaluate the output. Try adding more
specific phrases, examples or instructions to
personalize your results or remove things that you feel
are not necessary for you. You could try adding the
same instruction for detecting emoses like we
did in the last prompt, for example, by now, you have already
learned how to build prompts in both
markdown and XML, and you've got to
feel for the basics and the reasons behind
each part of the process. The next lesson, I will
show you how to set up proms that will help you automate what we
have been learning. So you can make
dictation templates for lots of different use cases. That will open up
a whole new set of possibilities for what you
can do with these tools. Stick with me, and I will
see you in the next class.
6. Specialized Dictation - Custom Use Cases: In this lesson, we're
going to dive into a more advanced level
of AI Power dictation. We will move beyond
general improvements to create truly specialized
and dynamic workflows. We're talking about
crafting prompts that together with
dictation can help you with text related
tasks that have specific requirements
beyond simple formatting. For this, we will start to
get into some features like context awareness that can make your dictation workflows
incredibly powerful. So far we have covered the basics of getting
a clean transcription. Then we learn how to
clean up that row text. In the last lesson,
I explained how to actually have some rewriting
rules and improve it. You also have learned how to
structure your problems with markdown and how to use XML for more complex
instructions. At this point, you have already
learned a lot, actually. Because of that, I want to share with you two prompts that will help you speed
up the process of coming up with your own
custom AI instructions. One is for creating simple
markdown based prompts. The other one is for generating more intricate XML structured before I walk you through
the rest of this lesson, I want to actually show you
how to get these prompts set up inside your
dictation app of choice. This is important because
when you start using AI for more assistant
related tasks, as I have mentioned before, not all of these apps
perform in the same way. In Superwhisper, it's
super straightforward. When you make a new custom mode, you can just drop the
full prompt writing it. Here, you are in full control. So whatever custom
prompt you want to use, just paste it and
you are good to go. Or spokene, the way to set this up is a bit hidden,
but it's still doable. When you create or edit
the prompting in Spokene, currently, you have to add
something in this space. I'm not sure why, and it
might change in the future, but it cannot stay empty. I will add a period.
Then head to the advanced settings and
find the system prompt area. That's the spot where you
paste your full custom. Spokenly and Superwhisper give you absolute control
over prompting, and this is something that is not yet available with voicing. I know this may
change in the future, but currently every time you
add a custom enhancement, you are limited to simple
text formatting tasks. I have also done tests in other dictation apps that follow your instructions,
but only partially. May need to do some testing
yourself with other tools. But one thing you can
also do is simply run these prompts with HGPT or
another AI service directly. Now that we have gone
through all the basics of building prompts and
experimenting on your own, I want to make sure
you get the most out of this prompt maker templates
that I'm sharing with you. The idea here is not to skip the learning you
have done so far. Already got a solid foundation,
so we can use that. Think of these prompt generators as shortcuts for
the heavy lifting, but it's still good idea
to slow down a bit and really tell the AI exactly
what you needed to do. So instead of giving the
prompt maker something like make me an AI that helps
me organize my thoughts, try getting a little bit more specific about
your workflow. For example, I am
already here within the interface when I am creating a new custom mode and
I can just dictate. I want a system prompt that takes a stream of
consciousness from the user, identifies the main ideas, pulls them out as bullet points, and then writes a short summary underneath that highlights
the most important points. Make sure that the
result also has a short but descriptive
title at the very top. The more clear and the more
detail your requirements, the more helpful the
resulting prompt will be. Perfect. I'm getting
all of this back. Now I will test it
in an empty node. First, I will switch to that mode that I
have just created. Right now I am using the new
prompt I have just created. I normally have an
instruction like this for recording thoughts
or ideas after reading something
or encountering a piece of content
I find interesting. I think something
like this is great for people who are into
knowledge management or note taking because they can find a piece of
information without having to focus on typing
or being slowed down by putting all of
their ideas in order, they can freely speak it and
get a result that is clearly organized and ready to be saved in a no taking up or
something similar. It's a great way to capture ideas that can later be
used for something else. Good. This looks much cleaner
than my original dictation. The thing here is that since you already know how
prompting works, if you see something
in the result that is not quite what
you were expecting, you can just go in there
and customize the prompting manually to make everything
fit your expectations. But the changes you will
need to do are minimal. You will not need
to start from zero. Now, maybe you can already start to see
the possibilities. With this, you are combining AI assistant related tasks
together with dictation. All of this starts to become even more powerful thanks to something that is called
context awareness. This is a feature that
allows you to send additional information
from your active window to AI whenever it is processing your transcription and
your instructions. Each dictation app handles
this a bit differently. You just have to know
the limitations of the tool that you are using
and work within that. For the apps that can detect your selected text, for example, you could dictate
something like, please make a list of the
tasks out of my selected text. Or you can also try. I need you to categorize
the different items in this list depending on the amount of work or
friction they require. It starts to sound
useful, right. You could also use this for reformatting something
you have already written, like selecting a
paragraph and telling the AI to put certain
words involved. You could ask for a
summary, translation. Yes, depending on the
app you are using, you can do all of that and more. The real power and time saving potential becomes
greater when you combine this feature with specialized
prompts that are unique to your own workflows and how
you like things to be done. Let me walk you through
a real world example. Let's say that you want to use dictation to answer
emails more efficiently. We will use the
prompt creator that will give me everything
with XML tags. This way, everything stays
clear and easy to follow. In Superwhisper, we already got one email formatting template, but let's do our own. I select my preferred models
and now let me dictate. I need your help,
creating an AI that can help me reply
to emails faster. I may provide the incoming
email or thread of messages as additional context
together with my response. Since this is meant for emails, please include a friendly
greeting at the top and add a sign of using my
name Robert at the end. I want the AI to understand the intention of my dictated
response or message, but organize it clearly, structure it appropriately,
and if necessary, elaborate on it to
give a better reply using additional context when available to improve the answer. Make it so that the AI only gives me the result
without any extra comments. Okay. Here goes my result. Do you remember that style guide that I mentioned in
the last lesson? I also paste that
here since I want the generated text to feel
less AI and more natural. I am using Superwhisper to activate those context
awareness features. I'll just select app context. I know this will grab the
content from my browser window, but I also have the option of clipboard context if I wanted. Now, let's test this. I'll go to an email where I received an offering to review
product and let me dictate. Thank you so much. Unfortunately, I don't
have time right now. Wonderful. The AI received the email that I
have on my screen. I understood what I
was trying to answer, and it elaborated a little bit in a way that still
feels natural. I would still go and do
a couple of quick edits, but this gets me so much closer to something that
I can quickly send out. As we wrap up this lesson
on custom use cases, here's what I want
you to try next. Pick the app that fits you best. Maybe Superwhisper spokenly. Ideally, something that allows you for both context awareness and prompt customization and spend a little time
exploring what's possible. A good chance for you to work on your project for the class. Find a use is that is unique
to you and try to come up with a solution with everything you have
learned so far. It could be anything,
for example, having a simple outline
in your front app, then dictating a stream
of consciousness based on that and having AI help you
organize everything nicely. That's something
that I often do or summarizing an article
in your front window, following a specific
set of guidelines. I don't know. I want you to
find something that will be genuinely useful and that
will boost your productivity. Right now we've got
these powerful apps that can help people in ways that were not possible before. And still a lot of
users that I've talked to only stick to the
most basic dictation. Now I'm giving you a good
excuse to go further than that. Don't be afraid to
tinker, run tests, and study the documentation
or the settings of your tool. Check how context awareness works in the specific
app you're using. Half the battle is
just figuring out what your tool is actually
capturing behind the scenes. And then it's just a
matter of thinking how you can use that
to your advantage. The moment you start connecting your own speaking habits
and creative needs with the apps features is when this whole process really
becomes like magic. Alright, stick around
for our last lesson. I have got a few more tips
that you will not want to miss as you keep
exploring dictation with AI.
7. Wrapping Up - Final Tips: Okay, guys, I am so happy that you have
completed this class. For me, all of this has been a real source of excitement
and constant learning. It was probably about
a year ago that I started experimenting
with these tools. Yeah, dictation has saved me an unbelievable
amount of time, and it has opened up
so many possibilities, especially as I have played with prompting,
context awareness, and combining it with
specific things I find myself doing day in and
day out around my system. Now, I want to give
you a quick recap, a few tips and thoughts
that might help you as you continue your journey
with this new generation of dictation tools. Now, first of all, I encourage you to keep
learning and experimenting, especially as new apps
and features pop up. Personally, I have been sticking to Superwhisper until now, but I'm always checking out
new options and workflows, just in case there's
something I can borrow or tweak to
find how I work. The way I see it, AI used in dictation should help
you express yourself. It should not drown out your voice or put you
on complete autopilot. There's something valuable
about deciding for yourself how much to
rely on speaking, typing, or even handwriting. None of these methods
need to disappear, and you don't have
to give up control. It's just about finding the
mix that lets you stay in charge and use these tools to truly support your
own creative flow. It's the same with prompts, by the way, there's a
place for simple prompts, and there's a place
for those more complex XML structures
that we talked about. This technology is moving
at an incredible pace, and I imagine that
prompting requirements or techniques will continue to
be simplified more and more. But even with that, with everything you have
learned in this class, you already got a very
strong foundation. In the end, what matters
most is that you know what you want and how
to get there yourself. I truly believe that
for those of you who have paid attention and put
in the effort to learn, you are going to have a real
advantage over other users who simply let AI take
every single decision. AI may have a lot
of training data and knowledge in so many fields, but it doesn't know about you or how you like
to get things done. If you can communicate clearly, which is what I have told you, you will be able to get much
more out of these tools. It's not just about
typing fastbord. It's about a huge
boost in productivity across so many different areas. I would really love for
all of you to head over to the project section here in Skillshare and share what
you have been working on. It would be awesome if you could share a prompt with everyone and tell us a bit about how you're using it in your dictation app. It's totally fine if
you are not using any of the advanced
features that I covered. Like context awareness. Even if it's a simple
text formatting prompt. If it's something you
have personalized and it works for you, it
would be great to see. On one hand, this would
let me know that you have got something out of the
class, but on the other hand, I think it can be useful
for everyone else to gather as inspiration or even to implement for
their own workflow. You have any questions
about anything we covered or if something
wasn't clear enough, we also have the discussion
section here on Skillshare. Feel free to pause in there. I don't know every single a
dictation tool out there, so I cannot give you very specific support on that aspect. But if I can help with anything
related to prompting or finding specific solutions
for one of your use cases, for example, I'll
be happy to do so. Finally, I would really
appreci if you could take a minute to leave a
review here on Skillshare. I would love to know
what you learned, what was the most useful and what you would like to
hear more about next. Even just a rating or
quick comment is great. It makes a big difference
for the visibility of the class and it helps
other students find it. If you are interested in
something along the same lines, I also have another class that
goes into writing with A. Where I focus more on
creative writing and share some of the
prompts that I use for brainstorming
wedding writing fiction. If that sounds like
something you would enjoy, definitely check it out. By the way, I also
run a YouTube channel where I cover more
advanced workflows, some of them related
to AI dictation and automation or other
productivity tools. If you would like to learn more, the channel link
is in my profile. Thank you so much for watching everyone. I'll see
you in the next one.
Robert J. P. Oberg, Creative • Filmmaker • Photographer
