Latest / AI Fire Daily / #623 Neil: ChatGPT Agent Workflows That Handle Repetitive Tasks Daily

Transcript

#623 Neil: ChatGPT Agent Workflows That Handle Repetitive Tasks Daily · AI Fire Daily

• 0:00
  There's this frustrating paradox with AI right
• 0:03
  now. We adopt these tools to save us precious
• 0:06
  time, but we end up spending hours copying prompts
• 0:09
  and moving data between apps. You're constantly
• 0:11
  managing the AI itself. You're still the busiest
• 0:14
  person in your own workflow. Exactly. I mean,
• 0:16
  we think we hired a brilliant robotic assistant.
• 0:19
  But we actually just bought a very demanding
• 0:22
  power tool. You still have to swing the hammer
• 0:24
  yourself every single day. Yeah, you really do.
• 0:27
  Welcome to the deep dive. Today, we are unpacking
• 0:29
  a fascinating guide on reclaiming your time.
• 0:32
  It shows how to stop managing ChatGPT by hand.
• 0:36
  We're going to explore building five autonomous
• 0:38
  agents. Right. And we will primarily use ChatGPT
• 0:41
  work for this. And for the heavier, complex reasoning
• 0:43
  tasks, we can leverage GPT -6 Astra. The ultimate
• 0:47
  goal is to handle your tedious workflows completely
• 0:49
  in the background. Okay, let's unpack this. First,
• 0:52
  we need to quickly define our terms. What exactly
• 0:55
  is an AI agent? Here's the simple version. An
• 0:57
  AI setup that runs specific tasks automatically
• 1:00
  on a schedule. That is exactly right. If we connect
• 1:04
  this to the bigger picture, the stakes here are
• 1:06
  quite high. This is about moving away from being
• 1:09
  an active participant in your software. You become
• 1:11
  the supervisor of a system that essentially runs
• 1:14
  itself. It's a massive identity shift. Oh, totally.
• 1:17
  You stop doing the digital chores and you start
• 1:19
  managing the actual outcomes. Because chat GPT
• 1:22
  work has persistent integrations with your enterprise
• 1:24
  workspace, it's not just a chat window anymore.
• 1:27
  Right. It becomes a true background process.
• 1:30
  Precisely. It operates while you are doing other
• 1:33
  things. Okay, so we've corralled our philosophy.
• 1:35
  Let's examine the first workflow. Agent number
• 1:38
  one is the commitment tracker. We're starting
• 1:40
  with the micro -level chaos of daily work. Because
• 1:43
  those tiny promises are always the easiest things
• 1:46
  to forget. Picture that classic Friday afternoon
• 1:49
  feeling. You told a client on Tuesday, I will
• 1:51
  send that file this afternoon. You promised a
• 1:54
  colleague in a hallway chat, I will follow up
• 1:56
  tomorrow. And you casually agreed to a random
• 1:59
  project deadline during a chaotic Zoom meeting.
• 2:02
  Yeah, and all those little commitments scatter
• 2:04
  everywhere across your digital life. They hide
• 2:06
  in your Gmail inbox. They get buried deep inside
• 2:09
  massive Slack threads. They sit entirely untouched
• 2:12
  in your daily meeting notes. And if you try to
• 2:15
  remember everything by yourself, you will eventually
• 2:17
  fail. A critical task will slip through the cracks.
• 2:20
  So the author uses the first chat GPT agent to
• 2:23
  solve this exact problem. It collects all these
• 2:26
  scattered commitments into one clean Notion database.
• 2:30
  It actively updates that database every single
• 2:33
  morning. The mechanism here is brilliant. You
• 2:35
  must connect your Gmail, your Slack, and your
• 2:38
  Meeting Notes app directly to chat GPT work.
• 2:41
  Wait, why not just connect your email? That seems
• 2:43
  vastly simpler to set up. Because many of our
• 2:45
  most important commitments happen outside of
• 2:47
  traditional email, if the agent only monitors
• 2:49
  your inbox, your database will miss critical
• 2:51
  tasks entirely. Oh, I see. Yeah. You need a comprehensive,
• 2:55
  overarching view of your communications. That
• 2:57
  makes total sense. Once the AI has that broad
• 3:00
  context, you give it a specific job. You instruct
• 3:03
  it to review those connected sources from the
• 3:05
  last seven days. Right. It specifically hunts
• 3:08
  for commitments, approaching deadlines, follow
• 3:10
  ups, or unassigned tasks. Then it creates a structured
• 3:14
  notion database called commitments and deadlines.
• 3:17
  It extracts highly specific fields for every
• 3:20
  single item it finds. It logs the task itself.
• 3:24
  It lists the counterparty involved. It notes
• 3:26
  the deadline. It links the original evidence
• 3:28
  and it assigns an owner. But what's fascinating
• 3:31
  here is a very specific rigid constraint the
• 3:35
  author adds to the prompt. Okay, what is it?
• 3:37
  The core instruction literally says if a deadline
• 3:40
  isn't clear stated, mark it as unclear. Ah, instead
• 3:44
  of guessing. That's a completely crucial addition
• 3:47
  to the prompt. Absolutely. AI systems naturally
• 3:50
  possess a strong desire to please the user. They
• 3:53
  desperately want to fill in all the blank spaces.
• 3:55
  I hate empty fields. Yeah. If you leave out that
• 3:57
  strict instruction, the AI might hallucinate
• 4:00
  a random deadline. Just wants to finish building
• 4:02
  the table. It's like stacking Lego blocks of
• 4:04
  data. Everything needs to click into the exact
• 4:06
  right place. If you don't give it strict rules,
• 4:09
  the AI tries to jam a square Lego piece into
• 4:12
  a round hole just to finish the build. It invents
• 4:14
  a date. You have to tell it, if the piece doesn't
• 4:17
  fit, leave it blank. That is a perfect analogy.
• 4:20
  You must force the AI to embrace uncertainty.
• 4:25
  That is how you avoid getting a long, hallucinated
• 4:28
  summary sitting inside your task manager. But
• 4:31
  wait, let me push back a little here. How do
• 4:33
  you actually know it worked correctly the very
• 4:36
  first time? Do you just trust the AI blindly
• 4:38
  with your entire to -do list? Not at all. You
• 4:41
  absolutely must review the first automated run
• 4:44
  manually. You check if it successfully found
• 4:46
  the right commitments. Makes sense. You verify
• 4:48
  that the evidence links actually match the original
• 4:51
  source documents. If the AI accidentally pulls
• 4:53
  in automated newsletters, you add a strict rule
• 4:55
  to ignore them. So you always check the AI has
• 4:57
  first run before automating it. Exactly. Once
• 5:00
  the output looks perfectly clean, you save it
• 5:02
  as a scheduled background task. It runs every
• 5:04
  morning while you drink your coffee. Okay, so
• 5:06
  we've corralled our own scattered promises, but
• 5:08
  that's just internal chaos. The real stress test
• 5:11
  for any system is when external chaos hits your
• 5:13
  inbox. Oh, yeah. That brings us to agent number
• 5:16
  two, the hiring dashboard. Tracking your own
• 5:19
  personal promises is one thing, but tracking
• 5:21
  hundreds of other people is a completely different
• 5:24
  monster. It gets incredibly messy. Imagine opening
• 5:27
  a job wreck and suddenly having 400 resumes to
• 5:30
  process. Job applicants pouring into a single
• 5:33
  inbox creates absolute chaos. It really does.
• 5:36
  You receive PDF resumes, short email intros,
• 5:39
  broken portfolio links, and endless follow -ups.
• 5:42
  And you usually have to manually migrate all
• 5:45
  of that unstructured data into a structured spreadsheet.
• 5:48
  I have to admit something here. I still wrestle
• 5:50
  with prompt drift myself when organizing massive
• 5:53
  email threads. It is mentally exhausting to keep
• 5:56
  the AI focused on extracting the right details.
• 5:58
  Well, this specific agent fixes that exhaustion.
• 6:00
  You feed ChatGPT work the actual hiring context
• 6:04
  first. By giving it a screenshot of the job posting,
• 6:07
  its vision model parses the visual hierarchy
• 6:10
  of the ad. It maps those requirements against
• 6:12
  the unstructured data in the applicant's email.
• 6:15
  Wait. Just a screenshot is actually enough for
• 6:18
  the AI? Yes. Chat GPT work can utilize that visual
• 6:22
  context directly without you typing anything.
• 6:25
  It understands the core role, the expected application
• 6:28
  source, and what specific applicant information
• 6:31
  actually matters. Then you give it a structured
• 6:33
  place to store all this messy data. You instruct
• 6:36
  it to automatically create a Notion database
• 6:38
  called Hiring Dashboard. It automatically generates
• 6:41
  specific fields, applicant name, current email,
• 6:44
  their current role, their past company, their
• 6:46
  design portfolio, and their LinkedIn profile.
• 6:48
  It automatically extracts their relevant industry
• 6:50
  experience and logs the application source. Crucially,
• 6:54
  it's their initial status field to new. It also
• 6:57
  follows a strict operational instruction to check
• 6:59
  the existing database first. It actively avoids
• 7:02
  creating duplicate applicant entries. That is
• 7:04
  highly intelligent. A candidate might send a
• 7:07
  second panicked email with a forgotten attachment.
• 7:10
  You certainly don't want them listed twice in
• 7:12
  your review pipeline. Right. Now here is where
• 7:15
  it gets incredibly powerful. The author utilizes
• 7:18
  a distinct two -step process for hiring. a second,
• 7:22
  entirely separate background workflow executes
• 7:26
  deeper candidate research. simply structures
• 7:29
  the messy data. The second agent actually analyzes
• 7:32
  that structured data. Exactly. The second workflow
• 7:34
  only looks at candidates possessing the new status
• 7:37
  tag. It independently researches them using their
• 7:39
  provided public links. Like their portfolio.
• 7:42
  Yeah. It scans their portfolio, their personal
• 7:44
  website, and their LinkedIn profile. It automatically
• 7:47
  appends a short research note to each candidate's
• 7:49
  notion record. It specifically focuses on highlighting
• 7:52
  past projects and skills directly connected to
• 7:55
  your job description. It actively hunts for useful
• 7:57
  qualitative signals for your upcoming review
• 8:00
  process. But there is a very strict non -negotiable
• 8:03
  rule applied right here. What's the rule? The
• 8:06
  AI is explicitly commanded, don't change the
• 8:08
  candidate status. Why limit the AI's capabilities
• 8:11
  like that? If it already knows their skills match
• 8:14
  the job, shouldn't it move good candidates to
• 8:17
  a reviewing status automatically? Because doing
• 8:19
  that completely removes the human from the final
• 8:22
  decision -making loop. This AI step should only
• 8:25
  add valuable context for your human review. Oh,
• 8:28
  I see. The AI is also instructed to clearly flag
• 8:31
  anything it cannot independently verify online.
• 8:35
  You absolutely don't want an AI secretly rejecting
• 8:38
  qualified people based on a hidden hallucination.
• 8:42
  Right. The AI adds context, but humans make the
• 8:45
  actual hiring decisions. Exactly. It empowers
• 8:47
  your final decision. It does not replace your
• 8:49
  judgment. Okay, so we've successfully structured
• 8:51
  our internal inbox. We have structured our external
• 8:54
  hiring pipeline. Now let's turn our attention
• 8:57
  outward to the internet. We're looking at agent
• 8:58
  number three. daily content ideas. Finding fresh
• 9:02
  content ideas sounds incredibly easy until you
• 9:04
  actually try doing it every single day. You open
• 9:07
  X, formerly known as Twitter. You instantly see
• 9:09
  hundreds of different posts saying the exact
• 9:11
  same generic thing. It is an incredibly noisy
• 9:13
  environment. Some posts gain real algorithmic
• 9:16
  traction because they are genuinely useful. Others
• 9:19
  just get inexplicably lucky with the timeline
• 9:22
  algorithm. This specific agent isolates the real
• 9:26
  traction. It actively filters out the weak, repetitive
• 9:29
  content. It permanently saves the genuinely useful
• 9:33
  ideas directly into Notion. First, you must give
• 9:35
  chat GPT work a highly clear research area. You
• 9:38
  define your core topics. For the author, it is
• 9:41
  tracking things like new AI agents, foundational
• 9:44
  models, and software product updates. You explicitly
• 9:47
  instruct the agent to focus purely on practical
• 9:49
  examples. Functional demos, strong confrarian
• 9:52
  opinions, major new software releases. And you
• 9:55
  explicitly command it to ignore general fluffy
• 9:58
  motivational posts. You actively filter out low
• 10:00
  -value repetitive reposts. You do not want random
• 10:03
  timeline posts that just happen to vaguely match
• 10:05
  a target keyword. Exactly. Once the primary research
• 10:08
  direction is securely set, you move the curated
• 10:10
  results into Notion. But there is a fascinating
• 10:13
  factual nuance buried in this step. The author
• 10:15
  explicitly does not use engagement metrics to
• 10:17
  filter the posts. This raises a profoundly important
• 10:20
  technical question. Why not simply filter the
• 10:22
  posts by likes or retweets? Well, the author
• 10:25
  notes that AI cannot always verify engagement
• 10:28
  data accurately. But why is that? Why does a
• 10:31
  smart AI fail at reading a simple number on a
• 10:33
  screen? It comes down to how these AI agents
• 10:36
  actually browse the web. They struggle heavily
• 10:39
  with dynamic JavaScript -heavy DOM elements.
• 10:42
  Ah, the code is too messy. Yeah. Things like
• 10:45
  Twitter's constantly updating engagement metrics
• 10:47
  are often hidden behind complex code or strict
• 10:49
  API rate limits. Scraping raw engagement data
• 10:53
  is highly unreliable for an LLM. So it just breaks.
• 10:56
  Pretty much. Therefore, prioritizing clear source
• 10:59
  quality is a vastly better filter than relying
• 11:01
  on easily manipulated likes. That makes perfect
• 11:04
  sense. So it builds a content research notion
• 11:06
  database based on quality, not just numbers.
• 11:09
  logs the post title, the authors, the exact URL,
• 11:12
  and the core topic. It pulls out the primary
• 11:14
  key takeaway. It also adds a very specific, brilliant
• 11:17
  field why it may be useful for content. That
• 11:20
  is incredibly clever. A source can be entirely
• 11:23
  new and completely accurate without actually
• 11:26
  being a good idea for a social media post. Exactly.
• 11:29
  It forces the AI to logically justify why it
• 11:32
  saved the link in the first place. Beat. Whoa.
• 11:35
  Imagine scaling that to a billion queries while
• 11:38
  you sleep. Having a tireless AI that logically
• 11:41
  justifies its own research choices to you every
• 11:44
  single morning. It is entirely possible with
• 11:46
  this exact architecture. You turn it into a scheduled
• 11:48
  daily agent. You run it every single morning
• 11:50
  at exactly 9 0 0 a .m. It only collects brand
• 11:53
  new items. It meticulously avoids creating duplicates.
• 11:57
  But what actually happens on a slow news day?
• 11:59
  What if there are literally no good updates in
• 12:02
  the AI space that specific morning? The author
• 12:05
  includes a strict non -negotiable rule for that
• 12:07
  scenario. If there are no strong new results,
• 12:10
  don't add weak content just to fill the database.
• 12:12
  Better to have an empty database than one filled
• 12:14
  with useless noise. Precisely. If nothing genuinely
• 12:17
  strong shows up, it simply leaves the database
• 12:19
  unchanged. We will be right back with the final
• 12:22
  two powerful agents after this short break. Stick
• 12:25
  around, sponsor Reid here. And we are back. OK,
• 12:29
  let's unpack agent number four, social drafts.
• 12:31
  We just built a sophisticated machine to find
• 12:34
  great ideas. Now we urgently need a machine that
• 12:37
  actually executes on those ideas. Right, because
• 12:40
  the daily research agent finds the compelling
• 12:42
  stories. But you do not want to open that database
• 12:45
  and stare blankly at a blinking cursor. You do
• 12:48
  not want to start writing from absolute zero.
• 12:51
  So we start chaining the automated workflows
• 12:53
  together. Agent number four automatically looks
• 12:56
  at the newest items generated by agent number
• 12:58
  three. It reviews the content research database.
• 13:01
  But it doesn't blindly turn every single research
• 13:03
  item into a finished post. It intelligently filters
• 13:06
  the research all over again. It asks itself a
• 13:09
  series of critical questions. Is the original
• 13:11
  story still recent? Is the underlying source
• 13:14
  actually clear? Is this topic genuinely relevant
• 13:17
  to my specific audience? It also asks, is there
• 13:20
  a highly useful angle here beyond simply repeating
• 13:23
  the break news, and it actively checks if the
• 13:25
  topic was already utilized in a previously published
• 13:28
  draft. It only keeps research items strong enough
• 13:30
  for actual content generation. Then it finally
• 13:33
  starts writing. But it follows incredibly strict
• 13:36
  stylistic guidelines. It writes remarkably simple,
• 13:39
  concise drafts. It crafts a clear, engaging opening
• 13:42
  hook. It explains the technical update in plain,
• 13:45
  simple language. It must keep the original source
• 13:48
  URL permanently attached to the draft. And crucially,
• 13:51
  it is explicitly told not to add random facts
• 13:54
  that are unsupported by the original research.
• 13:57
  It saves these generated texts into a brand new
• 14:00
  Notion database called social drafts. It marks
• 14:03
  their initial status simply as draft. Here is
• 14:06
  where it gets really interesting. There's an
• 14:08
  automated self -correction loop. built directly
• 14:11
  into this workflow. Yes, this is a phenomenal
• 14:13
  prompting technique. In AI, this is often called
• 14:16
  reflection. For an extra cleanup pass, you literally
• 14:19
  ask ChatGPT to critique and review its own generated
• 14:22
  drafts. It aggressively checks for repeated ideas.
• 14:25
  It actively looks for weak, boring openings.
• 14:28
  It hunts for unsupported claims or unnecessarily
• 14:31
  complex wording. It ruthlessly trims unnecessary
• 14:34
  length. It actively improves the draft's flow
• 14:37
  while meticulously keeping the original core
• 14:39
  meaning intact. But the author gives it one final
• 14:43
  absolute command. Don't publish anything. Why
• 14:47
  tell a highly capable autonomous agent to stop
• 14:50
  right before the finish line? Because the AI
• 14:53
  simply does not understand your current mood.
• 14:56
  It does not understand the delicate nuance of
• 14:58
  today's specific cultural context. Yeah, that
• 15:01
  makes sense. You still want to fix the final
• 15:03
  wording yourself. You want to inject your own
• 15:05
  personal context. You must remove anything that
• 15:07
  sounds slightly too generic before it goes public.
• 15:10
  The agent prepared the meal, but you decide when
• 15:12
  to serve it. Perfectly said. The automated workflow
• 15:15
  prepares the raw material. It does not make the
• 15:17
  final publishing decision for you. That perfectly
• 15:20
  leads us to our final workflow. Agent number
• 15:22
  five, the monthly business summary. We've successfully
• 15:26
  automated daily micro tasks. We have automated
• 15:29
  weekly content generation. Now we need to zoom
• 15:32
  out. We urgently need a macro -level view of
• 15:34
  the entire past month. The end of the month is
• 15:37
  always a chaotic blur. You have scattered Google
• 15:39
  Drive documents. You have random calendar events.
• 15:42
  You have endless overlapping Slack threads. You
• 15:45
  have critical files scattered everywhere. But
• 15:48
  you still need to accurately remember what actually
• 15:51
  happened. You need to know what truly moved the
• 15:54
  needle. This final agent automatically collects
• 15:56
  important context from the entire month. It distills
• 15:59
  it into one clear, readable business summary
• 16:03
  inside Notion. First, you must choose the integrated
• 16:06
  sources. Gmail, Slack, Notion, Calendar, Google
• 16:10
  Drive, and your meeting notes. But there is a
• 16:13
  vital non -negotiable safety check included here.
• 16:16
  Wait, I have to stop you there. Giving an AI
• 16:18
  free rein to read all my Slack threads and internal
• 16:20
  Google drives. Isn't there a massive privacy
• 16:23
  and security risk there? Girl, absolutely. How
• 16:25
  do we know it's not pulling highly sensitive
• 16:27
  HR data or private financial details into this
• 16:30
  summary? That is a completely valid concern.
• 16:32
  This is exactly why using enterprise level tools
• 16:35
  like ChatGPTwork is crucial. They have strict
• 16:37
  data governance. More importantly, before it
• 16:39
  generates the report, the agent must systematically
• 16:42
  confirm its own access permissions. It explicitly
• 16:46
  checks which specific sources and channels you
• 16:49
  have actually authorized it to read. Ah, so you
• 16:52
  wall off the sensitive data beforehand. If it
• 16:55
  cannot access a private channel, it must state
• 16:58
  that limitation clearly. It cannot just blindly
• 17:01
  guess what happened in a missing Slack conversation.
• 17:04
  Exactly. Once the authorized sources are confirmed,
• 17:07
  it reviews the month's data. It creates a comprehensive
• 17:10
  monthly business summary. It details important
• 17:13
  ongoing projects and measurable progress. It
• 17:15
  logs key strategic decisions. It summarizes major
• 17:18
  team conversations. It meticulously logs open
• 17:21
  commitments and unfinished work. It highlights
• 17:23
  important strategic wins. But how does this specific
• 17:26
  prompt prevent massive information overload?
• 17:29
  You don't want a literal transcript of every
• 17:32
  single email you sent for 30 straight days. Well,
• 17:35
  the prompt specifically asks the AI to identify
• 17:37
  problems or blockers that appeared more than
• 17:39
  once, and it specifically asks for items that
• 17:41
  demand your immediate attention next month. It
• 17:44
  filters out the noise to focus strictly on forward
• 17:46
  momentum. It highlights what actually matters
• 17:48
  for the future. It doesn't just blindly recap
• 17:51
  every tiny irrelevant task your team completed.
• 17:55
  And you schedule this comprehensive agent for
• 17:57
  the last day of every single month. It utilizes
• 18:00
  the exact same report structure every time, making
• 18:02
  it easy to compare months to sex silence. So
• 18:06
  what does this all mean? Let us slow down for
• 18:08
  a second and recap the big idea here. The core
• 18:11
  philosophy of this guide isn't actually just
• 18:13
  about speed. Speed is merely a helpful byproduct.
• 18:16
  The real lasting value is fundamentally shifting
• 18:20
  your identity within your own workflow. You're
• 18:22
  no longer an active stress participant frantically
• 18:25
  clicking around the app. You're the high -level
• 18:27
  orchestrator of the entire system. The AI finally
• 18:30
  takes care of the mundane work that constantly
• 18:33
  pulls you back into the same repetitive loops.
• 18:35
  A truly good workflow always starts with one
• 18:38
  clear, defined job. It needs enough background
• 18:41
  context, and it must reliably produce an output
• 18:44
  you can actually use. The author gives a very
• 18:47
  practical, grounded warning at the very end.
• 18:49
  Do not try to set up all five of these complex
• 18:51
  agents at once. That is a guaranteed recipe for
• 18:54
  intense frustration. You will get incredibly
• 18:57
  overwhelmed. The AI will inevitably make configuration
• 19:00
  mistakes. Start with just one single agent. Pick
• 19:03
  the workflow that successfully removes your most
• 19:05
  annoying, repetitive manual task. Verify that
• 19:09
  it works perfectly for a week. Then... Slowly
• 19:13
  expand from there. Build trust with the automated
• 19:16
  system slowly. Then confidently let it run. That
• 19:21
  is exactly when chat GPT work stops feeling like
• 19:23
  just another annoying browser tab you have to
• 19:26
  manage. It finally starts carrying the actual
• 19:28
  heavy workload. It finally becomes the autonomous
• 19:31
  digital assistant we were all promised in the
• 19:33
  first place. It shifts from being a tool you
• 19:35
  use to a system you manage. Exactly. I want to
• 19:37
  leave you with a final provocative thought to
• 19:39
  chew on today. If these autonomous agents are
• 19:41
  silently handling your inbox, your daily research,
• 19:44
  and your monthly business reviews. Yeah. What
• 19:46
  uniquely human skill should you be cultivating
• 19:49
  with all that newly free time? That is the real
• 19:52
  existential challenge of this entirely new era.
• 19:54
  Thank you for diving deep with us today. We will
• 19:56
  see you next time.