Latest / AI Fire Daily / #627 Neil: GPT-6 Astra Workflows That Actually Matter in ChatGPT Work

Transcript

#627 Neil: GPT-6 Astra Workflows That Actually Matter in ChatGPT Work · AI Fire Daily

0:00
Every time a new AI drops, they tell us it's
0:03
more powerful. But what happens when powerful
0:06
stops, meaning it gives better answers, and starts
0:08
meaning it does your actual job while you grab
0:11
a coffee? Right, yeah. I mean, that is the real
0:14
question today. It fundamentally changes the
0:16
entire way we interact with computers. Welcome
0:19
to this deep dive. Today, we are slowing down
0:22
a bit. We really want to unpack this highly practical
0:25
guide on GPT -6 Astra, specifically. We are looking
0:29
closely at a feature called Chat GPT Work. We
0:32
are going to cover setting up your workspace
0:34
properly. We'll look at auditing complex financial
0:37
files. Which is fascinating, by the way. Yeah,
0:39
totally. And we will talk about building reusable
0:42
logic frameworks or skills. We will also explore
0:45
coding desktop apps from scratch. Oh, and voice
0:48
delegation. We definitely can't forget that.
0:50
Right, voice delegation. And crucially, we will
0:53
cover managing your usage costs, cause that gets
0:56
expensive. Oh, yeah, it really does. Our thesis
0:59
of our source material here is this complete
1:01
paradigm shift. We are moving away from treating
1:03
AI like, you know, a simple search engine. Right,
1:06
like Google. Exactly. You do not just ask it
1:09
questions and wait for a paragraph anymore. You
1:11
really have to treat it like an asynchronous
1:13
workspace. You hand off a massive complex task,
1:18
you walk away, and then you just check the polished
1:20
results later. That feels like a massive behavioral
1:22
shift for most users, like really massive. Let
1:26
us start with the setup because the guide highlights
1:29
this critical difference right out of the gate.
1:31
Yeah. There is regular chat, which we all know.
1:35
And then there's chat GPT work. Right. So regular
1:38
chat is fundamentally designed for brainstorming.
1:40
It is this rapid back and forth conversation.
1:43
But the work interface operates on a totally
1:45
different premise entirely. You find it pinned
1:48
at the top of the desktop app. You select work
1:51
when you want the AI to actually execute a multi
1:53
-step task. You give it your raw notes. It creates
1:56
a dedicated file. And it edits that document
1:58
iteratively right there in the window. OK, let's
2:01
unpack this. Once you are in that work tab, you
2:04
actually have to actively select the model you
2:06
want to use. Yeah, this is super important. The
2:09
source heavily stresses choosing GPT -6 Astra
2:12
from the prop box, not just GPT -6. The names
2:16
look incredibly similar on the screen. It is
2:19
so easy to miss. It's a really common miss step.
2:22
I mean, if you pick the base model for a heavy
2:24
task, your complex workflow just fails to connect
2:27
the dots. The reasoning engine simply is not
2:29
deep enough. Yeah. The guide also mentions choosing
2:33
a thinking level for the AI. Let us define that
2:36
briefly. It is how much computing power the AI
2:39
uses to solve problems. Spot on. A higher thinking
2:42
level forces the model to spend way more time
2:45
reasoning before it outputs a single word, but
2:48
it drains your usage limits incredibly fast.
2:50
Right. You generally do not need the highest
2:53
level for daily tasks. A medium setting is usually
2:55
a perfect balance. You only push it to high for
2:58
deep forensic analysis. I kind of like to think
3:00
of it like giving an employee a dedicated desk
3:02
with files. You do not just chat with them at
3:04
the wall. cooler anymore. You give them a real
3:06
workspace, though I still wrestle with prompt
3:08
drift myself. You know, when the AI slowly forgets
3:11
your original rules over a long chat beat. Why
3:16
is it so critical to manually double check the
3:18
model before hitting enter on a big task? Oh,
3:22
because defaulting to a lesser model on a complex
3:24
workflow completely wastes your time. You end
3:27
up with incredibly shallow results. Oh, I see.
3:29
Yeah. It might summarize a few bullet points,
3:31
but it misses the nuanced connections. You basically
3:34
have to start over from scratch. So verifying
3:36
the model ensures you don't accidentally underpower
3:39
your complex multi -step workflows. You absolutely
3:42
need the right engine for the job. OK. So once
3:44
that workspace is set up correctly, we hit the
3:47
first true test of its reasoning capabilities.
3:50
finding hidden connections in messy financial
3:52
files. Yeah, this part is wild. The guide uses
3:55
this fascinating case study about a fictional
3:57
business, Northstar Studio. They basically just
4:00
dumped a single chaotic folder of data into the
4:03
workspace to see what the AI could make of it.
4:06
We are talking raw revenue reports, unpaid invoices,
4:10
scattered campaign data, vague project notes,
4:13
and random customer feedback all just sitting
4:16
in one digital pile. But the human user does
4:19
not just hit audit and hope for the best. That's
4:22
key. They run a very deliberate, crucial pre
4:25
-check prompt. Oh, right. They ask Astra to map
4:28
out the available information first before doing
4:30
any real analysis. The prompt explicitly tells
4:34
the AI If something is unclear, flag it instead
4:38
of guessing. I really love that approach. You
4:40
are treating it like a strict forensic accountant.
4:43
You are actively preventing AI hallucinations
4:45
by setting those strict boundaries. Yeah, so
4:47
the AI maps what specific files connect to what
4:50
overarching business areas. It spots missing
4:53
context or corrupted data before the real work
4:55
even begins. And once the human verifies that
4:57
initial map, the actual heavy audit starts. The
4:59
findings in the case study are incredible. Yeah.
5:02
I mean, Astra dug through the noise and found
5:03
a $12 ,600 overdue balance for a client called
5:07
Brightline. The brilliant part is how it connected
5:09
that unpaid invoice to a completely different
5:11
department. Wow. It noticed a new $9 ,800 order
5:16
for that exact same client, Brightline. And that
5:20
new order was already on the production floor.
5:22
Wait, so sales is booking a massive new job for
5:26
a client that accounts receivable has not even
5:29
collected from yet. Yeah. That happens all the
5:31
time in real businesses, but catching it requires
5:34
cross -referencing isolated departments. Astra
5:37
spotted that business risk instantly. It really
5:39
did. It also caught a duplicate $6 ,480 vendor
5:43
payment buried in the ledgers. It even flagged
5:46
a marketing campaign. The back -to -office search
5:49
campaign was bleeding money compared to the actual
5:52
leads it generated. The mechanism here is profound.
5:55
It did not just summarize individual spreadsheets.
5:58
It built a relational graph across different
6:00
business functions. It connected the sales team's
6:03
pipeline to the production floor to the accounts
6:05
receivable ledger. Looking at all those scattered
6:08
errors, unpaid invoices, bad marketing, duplicate
6:12
payments, my first reaction is just panic. Oh,
6:14
for sure. If an AI dumps all this on my desk
6:17
at 9 in the morning, I am completely overwhelmed.
6:20
Does the guide offer a way to actually manage
6:23
that chaotic output? It does. It introduces a
6:26
framework called the Fix First Action Plan. It
6:29
forces the AI to structure each urgent issue
6:32
logically rather than just listing complaints.
6:35
It breaks every finding down by the core problem.
6:39
the specific evidence found, the immediate business
6:42
risk, and the exact next action required by a
6:45
human. It transforms scattered data into a prioritized,
6:49
evidence -backed plan for immediate action. Exactly.
6:52
It removes the panic entirely. You know exactly
6:54
what fire to put up first. Okay, so the North
6:56
Star audit was brilliant. But if I have to write
6:59
a 400 -word prompt... Every single time I want
7:01
that level of forensic analysis, I'm going to
7:03
pull my hair out. Yeah, nobody wants to do that.
7:06
How does the guide suggest we automate the reasoning
7:08
itself? That brings us to building a skill. You
7:11
do not want to rewrite complex instructions daily.
7:14
You lock that logic into a reusable tool. Right.
7:17
The guide details creating something they call
7:19
a research -to -brief skill. The rules given
7:23
to ACTRA are incredibly strict. It must separate
7:26
proven facts from assumptions. It must actively
7:29
flag conflicting data. It must use simple, accessible
7:33
language. And it is explicitly forbidden from
7:36
adding unsupported claims. And so Astra internalizes
7:39
those rules into a repeatable, automated workflow.
7:43
Whenever you feed it new documents, it automatically
7:45
traps the evidence and connects the findings
7:47
based on those strict parameters. They gave this
7:50
specific skill a rigid five -part output structure.
7:53
First, the main question. Second, key findings.
7:56
Third, evidence. Fourth, risks or gaps. And finally,
8:00
recommended actions. Right. And the test case
8:02
for this skill involved a company called Aurelio
8:04
Workspace. The core question was simply what
8:07
they should prioritize in the next quarter. Astra
8:09
ingested the raw data and seamlessly linked cash
8:11
flow realities with fulfillment delays and marketing
8:14
metrics. It even intentionally leaves out sections
8:17
with no info. Right. And that prevents the generic
8:19
AI fluff we all hate. If there is no data to
8:22
support a section, it simply stays quiet instead
8:25
of hallucinating filler text. I am curious about
8:28
the formatting choice here. Why is forcing a
8:31
rigid, numbered output structure so vital for
8:35
this specific skill? It's all about predictability.
8:38
Consistency allows a human manager to review
8:41
complex decisions much faster. Oh, I see. Yeah,
8:44
when you look at the brief, you know exactly
8:45
where the risks and the hard evidence live every
8:48
single time. Rigid structures force the AI to
8:51
ditch fluff and only present proven, actionable
8:54
facts. Exactly. It elevates the output from a
8:57
neat summary to a truly functional business document.
9:00
Sponsor. We are going to take a very quick break
9:02
to hear from our partners. And we are back. We
9:04
have seen Astra process and organize text and
9:07
financial data perfectly. But if it can build
9:10
logical frameworks, can it build actual interactive
9:13
tools? Oh, absolutely. Like, can it manage things
9:15
in the physical world? The guide demonstrates
9:17
that it totally can. It walks through how to
9:20
build a fully functional desktop application
9:22
from scratch. They build an inventory restock
9:24
planner. I really love the methodology here.
9:27
It is built iteratively. It is like stacking
9:29
Lego blocks of data. Yeah. Great analogy. You
9:32
do not ask it to build the whole complex castle
9:34
at once. You ask for a simple foundation, you
9:37
test it, and once it holds weight, you ask it
9:39
to snap the next block on top. The core features
9:42
requested are highly practical for a physical
9:44
business. Simple SKU tracking, the ability to
9:48
import messy CSV files, and an algorithm calculating
9:51
exactly how many physical Unix need to be reordered
9:54
based on lead times. to sex silence. Handling
9:57
edge cases in physical inventory is notoriously
10:00
difficult. You have missing values in a supplier
10:03
spreadsheet. You have duplicate SKUs entered
10:06
by different employees. It's a nightmare. Astra
10:08
has to write code that accounts for all of those
10:10
messy, unpredictable human errors without crashing.
10:13
The user then asked Astra to improve the user
10:15
interface. They wanted low -stock items to be
10:18
visually easier to notice. They wanted to sort
10:20
the urgent items by priority automatically. They
10:23
essentially asked to make the raw data table
10:25
scannable for a tired warehouse manager. And
10:28
Astra implemented all those visual changes, but
10:30
the impressive part is the validation. After
10:33
altering the code, Astra ran and passed all 16
10:36
automated tests. The core mathematical features
10:39
did not break when the visuals changed. I want
10:42
to pause on that mechanism. How does Astra handle
10:45
that testing phase without the user having to
10:48
manually write Python code to fix the inevitable
10:50
bugs? The AI runs a loop. It detects its own
10:53
bugs by reading the error logs, rewrites its
10:56
own code to fix them, and reruns the test suite
10:58
automatically until the main workflow functions
11:01
perfectly. It acts as both the lead developer
11:04
and the QA tester in one loop. Yep. The human
11:07
is just providing the product vision. The AI
11:09
manages the entire development cycle. If you
11:11
were listening to this and realizing you could
11:13
have Astra running complex audits, building skills,
11:16
and coding apps all at once, you also realize
11:18
typing out instructions for each one becomes
11:20
a massive bottleneck. Oh, yeah. It takes forever.
11:23
How do you manage that level of chaos? You bypass
11:26
the keyboard entirely. You move to live voice.
11:29
This is where the asynchronous nature of the
11:30
workspace really truly shines. You can give multiple
11:34
completely unrelated instructions in a single
11:37
continuous breath. The guide shows a fascinating
11:40
dual prompt. Yeah, so the user opens the live
11:43
voice feature on their phone. First, they casually
11:45
ask the AI to update a supplier shortlist. The
11:49
parameters are strict, though. Only keep suppliers
11:51
with a unit cost under $18 and a shipping lead
11:54
time under 21 days. Second, without pausing,
11:58
they say to rewrite an upcoming product launch
12:01
email. They want the copy to focus heavily on
12:03
product durability, and they want the previous
12:06
discount angle removed entirely. Those are two
12:08
completely different intellectual contexts, two
12:11
different isolated projects running silently
12:13
in the background of the workspace. Whoa, Pete.
12:16
Imagine it managing dozens of complex background
12:18
tasks at once, just from a 30 -second voicemail
12:21
while you walk to your car after a meeting. I
12:23
know, right? It acts as the ultimate executive
12:25
assistant. It takes a messy stream of consciousness
12:27
and delegates it perfectly. But how does the
12:30
model avoid bleeding information across tasks?
12:32
Like, how does it not mix up the supplier costs
12:35
with the email copy? Astra utilizes deep contextual
12:39
mapping. It parses the spoken request, separates
12:42
the intents, and routes each instruction to the
12:45
specific isolated project workspaces running
12:48
asynchronously. It routes your voice commands
12:50
perfectly to distinct isolated project workspaces
12:53
without confusion. Exactly. It parses human intent
12:57
beautifully, even when we speak in rambling run
13:01
-on sentences. This level of heavy asynchronous
13:04
multitasking sounds amazing. but it obviously
13:07
requires immense computational power. Oh, ton
13:09
of it. If you use Astra for everything, your
13:12
monthly usage limits will vanish in a day. What
13:15
is the strategy to prevent that? You have to
13:16
be highly deliberate about resource allocation.
13:19
The guide emphasizes a strict sorting of tasks
13:21
between the heavy GPT -6 Astra and the lighter
13:24
GPT -5 .6 model. So Astra is strictly reserved
13:27
for deep reasoning. the messy complex stuff.
13:30
Financial file audits, writing iterative code,
13:32
and processing that chaotic live voice delegation.
13:35
Right. You do not use Astra for regular predictable
13:37
email checks. You do not use it for daily news
13:39
monitoring or like simple summaries. You assign
13:41
those routine lightweight tasks to GPT 5 .6.
13:45
The author of the guide has an interesting daily
13:47
methodology. They test a new workflow prompt
13:50
on both models simultaneously first. If the outputs
13:54
are reasonably close in quality, they delegate
13:56
that task permanently to 5 .6. And that singular
14:00
habit saves your astral limits for when it actually
14:03
matters, for when you desperately need that deep,
14:06
connective reasoning to solve a hard problem.
14:09
When you are scaring at a new daily workflow.
14:12
What is the ultimate deciding factor when assigning
14:15
it to a specific model? You just analyze the
14:18
complexity of the reasoning required. If the
14:20
task is simply extracting basic text from a document,
14:24
deploying the heavy model is completely wasted
14:26
compute. Always test with the cheaper model first
14:28
to protect your heavy compute usage limits. Yeah,
14:31
you really have to treat your computing power
14:32
like a strict financial budget. This has been
14:35
a really fascinating journey into how we work.
14:37
Let us recap the big idea we explored today.
14:40
We have to stop using ChatJPT as just a super
14:43
-powered search engine that answers trivia. Exactly.
14:46
We need to start treating it as a true asynchronous
14:49
workspace. Real complex work is delegated, executed
14:53
in the background, and then reviewed by you.
14:56
You are the manager. The AI is doing the connecting.
14:59
I highly encourage you to check your own workflows
15:01
this week. Look closely at where you are wasting
15:03
hours due to tedious connecting yourself. There
15:06
is almost certainly a task you can hand off to
15:08
a skill right now. Think about that one messy
15:11
unstructured project sitting on your desk right
15:14
now. The one you've been avoiding for weeks.
15:16
Beat. What would happen if you dropped all those
15:19
files into a folder today and let Astra find
15:22
the hidden connections you missed? It might just
15:24
solve your biggest professional headache while
15:25
you grab a coffee. Thank you for joining us on
15:27
this deep dive. Stay curious.