Latest / AI Fire Daily / #649 Neil: Gemini Skills Turn Repeated Tasks Into Reusable Workflows

Transcript

#649 Neil: Gemini Skills Turn Repeated Tasks Into Reusable Workflows · AI Fire Daily

0:00
There's this sheer repetitive exhaustion that
0:03
we face trying to use AI at work. Just copy pasting
0:07
the exact same instructions and rules into the
0:09
prompt box every single morning. Right. And then
0:13
finally fixing it because it forgot half of what
0:15
you said anyway. I mean, treating your AI like
0:19
an amnesiac every day is just a terrible way
0:22
to work. Welcome back to the deep dive. If you
0:24
are tired of retraining your AI from scratch
0:26
every time you open your laptop, well, Today
0:30
is going to change your relationship with these
0:31
tools. Yeah, we are looking at Gemini skills
0:34
today. Right. Our roadmap for this conversation
0:36
is pretty straightforward, but the implications
0:38
are massive. We will unpack what these skills
0:42
actually are and the three different ways you
0:44
can build them. And we're going to talk about
0:46
how to use and manage them without creating a
0:48
complete digital mess. Exactly. And finally,
0:51
we'll explore how to combine them. It's like
0:53
stacking Lego blocks of data to automate complex
0:56
multi -step workflows. I'm really energized by
0:59
this, actually. It fundamentally shifts the dynamic.
1:02
You move from being an exhausted prompt writer
1:04
to, you know, essentially managing a system that
1:07
works for you. Yeah, we spend a staggering amount
1:09
of time repeating ourselves. We explain our business
1:12
context, our tone, our formats just over and
1:15
over. Let's solve that problem today. So let's
1:18
start with the foundation. What exactly are these
1:22
Gemini skills? Well, to appreciate the solution.
1:25
We have to really understand the core problem
1:27
with standard AI chat interfaces. Whenever you
1:30
open a new chat, the AI starts with a completely
1:33
blank slate. Right. It's a brand new relationship
1:36
every single time. It knows absolutely nothing
1:38
about your specific business needs or clients.
1:41
It doesn't know your internal jargon or your
1:44
style choices. So you are forced to explain the
1:46
entire job from scratch. You tell it what role
1:49
to play. You tell it what input you are providing.
1:52
And if you miss even one detail, the whole output
1:55
is skewed. You're essentially onboarding a new
1:58
employee every morning. That's exactly what we
2:01
want to avoid. So Gemini Skills fix this friction
2:05
by giving you reusable workflows right inside
2:07
Gemini Chat and Spark. Yeah. And just to clarify
2:11
our terms, a workflow is simply a repeatable
2:14
series of steps to finish a task. Exactly. These
2:17
skills basically allow you to lock in your custom
2:20
instructions. Permanently. It is a dedicated
2:23
workspace. Right, where your specific rules,
2:26
your preferred tone, and your background context
2:28
are saved and always active. And goes beyond
2:31
just text instructions, doesn't it? Oh, it does.
2:34
This is where it gets incredibly powerful. You
2:36
can attach supporting files directly to the underlying
2:39
skill. Like uploading a PDF -style guide for
2:42
reference. Exactly. Or you might upload a dense
2:45
CSV file full of your current quarterly data.
2:48
So the skill remembers your rules perfectly.
2:51
It remembers your exact formatting preferences
2:53
every single time you invoke it. Yeah, it gives
2:55
the AI all the specialized knowledge it needs
2:58
up front. It learns the constraints before it
3:00
even looks at the specific daily task. It removes
3:03
that exhausting friction from your work process
3:05
entirely. So the skill essentially acts as a
3:08
permanent specialized memory bank. Yes, it holds
3:11
your exact parameters and context locked in.
3:14
Got it. Permanent memory, zero daily retraining
3:18
required. Precisely. Okay, writing a giant prompt
3:21
makes sense if you are starting from scratch,
3:23
but how do we actually construct a skill without
3:26
getting overwhelmed? Well, there are three main
3:28
ways to create these skills. We should walk through
3:31
each method because they serve different needs.
3:33
Right. If your workflow is relatively simple,
3:36
setting it up with a direct prompt is usually
3:38
the fastest route. Let's unpack that first method.
3:40
Building a skill with a prompt. You basically
3:43
describe the entire architecture of the job in
3:45
one go. Yeah, you describe the role, the expected
3:48
input, the exact step -by -step process, and
3:50
the rigid output structure. Take a meeting summary
3:53
skill, for example. Right, because raw meeting
3:55
transcripts are usually a chaotic mess of crosstalk
3:58
and half -finished thoughts. Exactly. So you
4:01
tell the AI its purpose, which is turning those
4:03
raw, messy notes into a crystal clear summary.
4:06
Then you set very strict, non -negotiable rules
4:08
for the process. You instruct it to extract the
4:11
main topics. But you specifically tell it to
4:13
separate confirmed decisions from random brainstorming
4:16
ideas. You also tell it to extract action items
4:19
only if there is a clear assigned task. Yes.
4:22
And you can tell it to merge repeated points.
4:25
You tell it to strip out all the casual small
4:27
talk. But perhaps most importantly, you explicitly
4:30
tell it what not to do. Like, you tell it not
4:33
to guess missing names or dates. That is crucial.
4:37
Large language models are prediction engines,
4:39
you know? They want to fill in the blanks. So
4:41
by building a skill, you can hard code a rule
4:43
that says, do not invent a deadline if one isn't
4:47
mentioned. Right. Giving the full job architecture
4:49
upfront makes the skill incredibly reliable.
4:52
But what if I have massive amounts of data? Yeah,
4:56
you don't want to type all that out into a prompt
4:57
box. That brings us to the second method, which
5:00
involves uploading a file directly. You might
5:02
already have a markdown file containing all your
5:05
complex instructions, a skill .md file. You just
5:08
upload this directly into your workspace alongside
5:11
any supporting data files. Give us a concrete
5:14
example of why you would use this. Think about
5:16
building a proposal writer skill for a sales
5:18
team. You upload your markdown file with the
5:21
writing instructions, the tone, the layout. But
5:24
you also attach a CSV file containing your actual
5:28
up -to -date pricing data. So the AI uses that
5:31
CSV file as a fixed reference point. Yes, it
5:34
acts as an anchor. The AI depends on that fixed
5:37
information to write the proposal. This fundamentally
5:41
stops the AI from hallucinating incorrect numbers.
5:44
I won't invent a 20 % discount just to make the
5:47
text sound more persuasive because it is bound
5:50
by the uploaded file. Exactly. I have to admit,
5:53
I still wrestle with prompt drift myself. I find
5:56
myself typing out these massive lists of instructions
5:59
for complex editing tasks. And it just loses
6:01
track. Exactly. The AI just forgets the rules
6:04
halfway through a long conversation. It is incredibly
6:07
frustrating. You aren't alone. Manual typing
6:09
over long sessions often pushes the earliest
6:12
instructions out of the AI's active memory window.
6:15
That's why building a rigid skill is better,
6:17
but... There is a third creation method that
6:20
actually solves that specific frustration perfectly,
6:23
right? Oh, absolutely. The third method is turning
6:26
a successful conversation into a skill automatically.
6:29
Let's say you just spent an hour wrestling with
6:31
the AI. You refined a workflow over 20 back and
6:34
forth messages. Yeah, correcting its mistakes,
6:37
adjusting the tone, until the output was finally
6:39
perfect. You got the perfect result, but the
6:43
process was messy. Right. Instead of trying to
6:47
manually reverse engineer how you got there,
6:49
you just ask Gemini to do it. You ask it to review
6:52
the entire chat history. You tell it to turn
6:55
the final successful workflow into a reusable
6:58
skill. It extracts the rules and output format
7:00
for you. It does. And brilliantly, it ditches
7:04
all the failed attempts from the chat history.
7:06
It strips away your temporary corrections. It
7:09
just saves the polished logic that actually worked.
7:12
You are preserving your success. Yeah, it's amazing.
7:15
Which of these three creation methods is the
7:18
most foolproof for beginners? Turning a successful
7:21
pass conversation into a skill is absolutely
7:24
the best because it is already field tested in
7:27
reality. Makes sense. Extracting rules from successful
7:30
chats minimizes errors. It really is the most
7:33
practical way to start. So we have built this
7:35
highly specialized, perfectly instructed veteran
7:38
assistant. Let's talk about how we actually deploy
7:41
it when we are in the middle of a busy workday.
7:43
Using these skills in your daily routine is remarkably
7:46
simple. It is designed to keep you in your flow
7:49
state. You don't have to navigate a way through
7:51
complex menus. Not at all. How does the activation
7:53
actually work? You just type a slash. Just a
7:56
forward slash? That's it. Right in the normal
7:57
window. Yes. A menu pops up. You select your
8:01
custom skill. You paste your raw input. And hit
8:04
enter. The response is almost instant. The complex
8:07
instructions run silently in the background.
8:10
Applying your formatting and tone automatically.
8:13
Without any extra typing. You just check the
8:15
output. The system strictly follows the pinned
8:17
context. Does using these skills disrupt your
8:21
normal chatting flow in the interface? No, the
8:23
slash command keeps everything seamlessly in
8:25
the exact same chat window. Typing a slash keeps
8:28
everything seamlessly in one window. It's brilliantly
8:31
simple. Sponsor. Welcome back. Before the break,
8:35
we gave the AI a perfect permanent memory for
8:38
one specific job. We learned how to activate
8:40
a single skill seamlessly, but real work rarely
8:45
happens at a vacuum, right? Yeah, it requires
8:48
passing information through different stages
8:50
of refinement. This is where we transition from
8:53
a simple shortcut to a powerful automated system.
8:57
This is where things get truly fascinating. We're
9:00
talking about chaining them together. We're talking
9:03
about stacking those Lego blocks of data to automate
9:06
complex multi -step workflows. Splitting complex
9:10
tasks into smaller specialized skills makes the
9:12
entire process much more reliable. If you give
9:15
an AI a prompt with 50 different rules for writing
9:18
and editing, it is going to drop the ball. But
9:21
if you split them into a writer skill and a reviewer
9:24
skill, the accuracy skyrockets. Let's walk through
9:28
that specific digital assembly line. It's like
9:30
having station one bolt the chassis together
9:32
and station two hit it with a hammer to see if
9:35
the claims hold up. That is a great analogy.
9:38
First, you invoke the proposal writer skill.
9:41
You feed it a messy client brief. The writer
9:44
drafts the initial proposal, pulling verified
9:46
information from your attached pricing sheets
9:48
and service catalogs. So it generates a solid
9:50
initial draft automatically, right? Then you
9:53
take that draft and hand it to the proposal reviewer
9:56
skill. This second skill acts as your ruthless
9:59
automated quality control department. What exactly
10:02
is this automated manager checking for? Well,
10:04
it checks the mathematical calculations. to ensure
10:07
the pricing matches the uploaded CSV. It verifies
10:10
the scope of work against your standard deliverables.
10:13
And most importantly, it specifically hunts for
10:15
any unsupported claims or strange guarantees.
10:18
Because sales teams often promise things by accident
10:21
in their raw notes, like we offer 24 -7 support
10:24
when the official policy is strictly nine to
10:27
five. Exactly. The reviewer skill catches those
10:30
dangerous, expensive promises before the document
10:33
ever reaches the client. It acts as a digital
10:36
safety net. It also flags missing information
10:38
that definitely needs human confirmation. Whoa,
10:42
imagine setting up entire digital assembly lines
10:45
like this. You could scale your quality control
10:47
across a massive organization effortlessly. It
10:50
is a profound shift in how we approach knowledge
10:53
work. The AI is literally checking its own work,
10:56
policing its own hallucinations based on your
10:59
strict uploaded rules. And crucially, if the
11:03
reviewer finds an issue, It doesn't just give
11:06
up. No, it reacts. It doesn't just passively
11:09
point out the error for you to fix manually.
11:11
It triggers a revision loop automatically. It
11:14
sends the flag document back, revises the proposal
11:18
to fix the error, and runs the entire review
11:20
process again. It keeps looping until the document
11:23
meets every single established criterion. Once
11:26
it passes, it outputs three specific things for
11:29
you. First, the final polished proposal document.
11:33
Second, a brief summary of the review process
11:35
it conducted and what it changed. And third,
11:39
a final status update. The status is either ready
11:42
to send or needs human review. Because again,
11:45
you have strictly told it not to invent missing
11:48
facts. If the reference files don't resolve an
11:50
issue, it marks it clearly for you. It refuses
11:53
to guess. That discipline is the most valuable
11:56
part of the reviewer skill. Can the reviewer
11:59
skill actually force a rewrite without my intervention?
12:02
Yes. The prompt instructs it to revise and run
12:04
the review again automatically if an issue is
12:07
found. Incredible. It triggers an automatic revision
12:10
loop without me. It really is a game changer
12:13
for solo operators or small teams. But with all
12:16
these automated systems running, we have to talk
12:18
about maintenance. Yet we need to discuss digital
12:21
housekeeping, or else this powerful system is
12:23
going to spiral out of control. It is a very
12:26
real risk. If you aren't careful, you end up
12:29
with a cluttered workspace full of outdated instructions.
12:31
So let's talk about managing these tools. You
12:34
handle this through your workspace settings.
12:36
Yes. Next to any active skill, you have a management
12:39
menu. You can manually tweak the instructions
12:42
if your process changes slightly. You can ask
12:44
the AI to help edit the skill. You can deactivate
12:47
a skill you aren't using right now without permanently
12:49
deleting it. But the most powerful management
12:52
feature is the ability to replace the supporting
12:54
files, right? By far. That feature is critical
12:57
for long -term workflows. Walk us through why
13:00
replacing a file is better than just editing
13:03
the prompt. Well, imagine it is the end of Q3.
13:06
Your pricing CSV has changed because you have
13:09
raised your rates for Q4. Instead of opening
13:12
a massive instructional prompt and trying to
13:14
manually find and update every single price point
13:17
in the text, you simply swap the file. You use
13:20
the replace skill option, delete the old CSV,
13:23
and upload the new Q4 pricing spreadsheet. Exactly.
13:26
The core instructional logic of the skill remains
13:28
untouched. But the next time you type that slash
13:31
command, it immediately starts using the new
13:33
data. It is a localized update that saves you
13:36
hours of rebuilding. But there is a clear warning
13:41
that comes with this level of easy customization.
13:44
You need to keep your active skills list deliberately
13:46
lean. You should not create a new skill for every
13:49
tiny insignificant task you do once a month.
13:52
Because if you have 75 different active skills,
13:57
finding the right tool in that slash menu becomes
13:59
incredibly annoying. It defeats the entire purpose.
14:03
The goal is removing friction, not replacing
14:05
prompt friction with menu friction. What is the
14:08
biggest risk if I just start making skills for
14:10
absolutely everything? A cluttered skill list
14:13
makes finding the right tool frustrating, totally
14:16
defeating the purpose of the feature. Right.
14:18
Too many skills just recreates the original clutter.
14:21
We really need a practical philosophy for when
14:24
to lean on these tools and when to just use standard
14:26
chat. Let's synthesize all this. We need a clear
14:29
framework for when to use a skill and when to
14:31
avoid it. The ideal use cases fall into a few
14:35
distinct buckets. First, tasks that repeat frequently
14:38
and follow a clear, established process. If you
14:41
do it the exact same way twice a week, it is
14:43
a candidate for a skill. Structured writing is
14:46
a perfect example. End -of -week marketing reports
14:49
or client onboarding emails that always demand
14:51
a fixed output format. The AI follows your rigid
14:55
structure across different weekly inputs. Working
14:57
with fixed reference files is the second bucket.
15:00
Anytime you need the AI to reference a CSV or
15:04
a dense legal PDF. Or a brand style guide. Instead
15:08
of relying on its general knowledge, you should
15:10
build a skill. It anchors the AI's predictions
15:12
to your reality. Multistep workflows, like our
15:16
writer -reviewer assembly line, are the third
15:19
bucket, separating the creation step from the
15:22
quality control step. And finally, preserving
15:25
refined processes. When you have spent an hour
15:28
wrestling with a complex task and finally nailed
15:31
the logic, save it as a skill. You are capturing
15:33
lightning in a bottle so you never have to rebuild
15:36
that specific workflow again. We must contrast
15:39
this with tasks that don't fit the model at all,
15:41
though. If a task changes its fundamental nature
15:44
every single time you do it, it does not need
15:47
a skill. You shouldn't force a dynamic fluid
15:50
problem into a rigid structured system. If you
15:52
can't clearly define the rules and the desired
15:56
output, the skill is going to fail anyway. Should
15:58
I be using this for creative, unpredictable brainstorming
16:02
sessions? No. Skills are for fixed structured
16:06
processes, not one -off fluid ideation where
16:09
the goal is exploration. Okay, stick to structured
16:12
tasks. Avoid using for pure ideation. Keep the
16:16
creative wandering in the standard chat window.
16:19
Let's bring this entire deep dive full circle.
16:22
The ultimate paradigm shift here isn't just about
16:24
saving a few keystrokes. It is about changing
16:27
how we view our daily interaction with computers
16:29
entirely. You have to stop treating every repeated
16:32
task like a brand new conversation. It truly
16:35
alters the dynamic. Once you lock in your rules,
16:38
your reference files, and your required structure,
16:41
the friction of repeated work vanishes. You keep
16:44
your output perfectly consistent without having
16:46
to micromanage the AI every single morning. So
16:49
here is a piece of actionable advice to take
16:51
with you today. Look at your calendar. Pick just
16:55
one tedious structured task you already repeat
16:58
every single week. Turn it into a skill today.
17:01
Run it on real work and see if it saves you from
17:03
having to explain yourself tomorrow. Just one
17:06
task is all it takes to see the value. If we
17:08
can now easily save, combine, and deploy the
17:11
absolute best versions of our workflows to run
17:13
silently in the background, well, how long until
17:16
we are entirely managing a team of our own customized
17:20
digital experts, rather than doing the daily
17:23
drafting ourselves? Thank you for joining us
17:25
on this deep dive.