Latest / AI Fire Daily / #3 Eric: Master the gemini AI: Messy brief to final result

Transcript

#3 Eric: Master the gemini AI: Messy brief to final result · AI Fire Daily

• 0:00
  Imagine asking an AI for a brand new campaign
• 0:03
  plan. Right. And it hands you an absolute masterpiece
• 0:06
  of a document. But it built the whole thing on
• 0:08
  hidden assumptions. Yeah. I mean, a polished
• 0:11
  answer can definitely still be the wrong answer.
• 0:13
  Welcome to the Deep Dive. Today, we are looking
• 0:15
  at how to master the Gemini app. We are learning
• 0:18
  how to turn a vague task into really clear output.
• 0:23
  We want to teach you how to maintain total control.
• 0:26
  Exactly. So the AI doesn't make crucial business
• 0:28
  decisions for you. Before we can control this
• 0:31
  tool, well, we have to understand it. We need
• 0:34
  to see why it makes mistakes so confidently.
• 0:37
  We usually call this phenomenon the Gemini trap.
• 0:39
  Say you give the AI a very vague prompt. Like,
• 0:42
  create a campaign plan for an AI webinar. Right.
• 0:46
  Gemini goes to work immediately on that vague
• 0:48
  request. It creates a highly professional, well
• 0:51
  -structured campaign plan. It looks completely
• 0:54
  finished. But it secretly assumes who your target
• 0:56
  audience actually is. It assumes the core offer
• 0:59
  of the webinar, too. Yeah. And it completely
• 1:01
  invents the campaign goals and the success metrics.
• 1:04
  It is just full of unverified, hidden assumptions.
• 1:08
  Right. And it looks absolutely beautiful on the
• 1:10
  screen. It uses bold headers and perfect little
• 1:13
  bullet points. So you are incredibly tempted
• 1:16
  to just accept it. We are wired to trust things
• 1:19
  that look highly professional. To avoid this
• 1:22
  trap, you know, we have to look under the hood.
• 1:24
  We need to understand the fundamental mechanics
• 1:26
  at play here. Yeah. Gemini doesn't actually think
• 1:29
  like a human being does. No, it doesn't pull
• 1:31
  guaranteed facts from a rigid, fixed database.
• 1:35
  Instead, the system simply processes your current
• 1:38
  conversational context. It relies heavily on
• 1:41
  massive sets of previously learned patterns.
• 1:44
  Then it combines those recognized patterns with
• 1:46
  your specific prompt. It looks very closely at
• 1:48
  your recent chat history. It also scans any files
• 1:51
  you upload to the window. It uses what we call
• 1:53
  multimodal generative models. Meaning what, exactly?
• 1:56
  That is AI processing text, images, and audio
• 1:59
  to create content. It stitches all these different
• 2:02
  data formats into one context. Beat, that makes
• 2:05
  a lot of sense when you break it down. It is
• 2:07
  basically predicting the most probable next piece
• 2:09
  of information. Yeah, it is kind of like an eager
• 2:12
  intern trying to impress you. They fill in the
• 2:14
  blanks without asking any clarifying questions.
• 2:17
  They just desperately want to please you with
• 2:19
  a finished product. Beat. I still wrestle with
• 2:22
  prompt drift myself, honestly. No, for sure.
• 2:24
  Long context windows are tricky for anyone to
• 2:27
  manage properly. The machine simply loses focus
• 2:30
  over time. Your core instructions just get lost
• 2:33
  in all that text. The AI completely loses the
• 2:37
  thread over a long conversation. So it's prioritizing
• 2:40
  learned patterns over the absolute truth. Yes,
• 2:43
  the text generation process is entirely probabilistic
• 2:46
  by nature. It predicts the most likely next word
• 2:48
  based on the input. It doesn't know facts. It
• 2:51
  just knows mathematical probability. Got it.
• 2:54
  It gives the most likely answer, not the truest.
• 2:56
  Exactly. It just wants to complete the pattern
• 2:58
  you started. If you start a corporate plan, well,
• 3:00
  it finishes a corporate plan. Since the AI naturally
• 3:03
  wants to guess and please us, how do we establish
• 3:07
  ground rules to actually stop it. You have to
• 3:09
  step up and take firm control immediately. The
• 3:12
  human user has three specific non -negotiable
• 3:15
  responsibilities here. Okay, let's unpack those
• 3:17
  three distinct responsibilities. First is context.
• 3:21
  Right, you must control the context. Second,
• 3:24
  you must strictly control the decisions. And
• 3:27
  third, you must control the verification process.
• 3:30
  Context is the actual information the AI is allowed
• 3:32
  to use. Yeah, you are fencing in the data it
• 3:35
  can pull from. Then, decisions are the strategic
• 3:39
  choices the human explicitly makes. You cannot
• 3:42
  outsource your strategy to a math equation. And
• 3:45
  finally, verification is the criteria you use
• 3:47
  for success. You define what a good outcome looks
• 3:50
  like. So we need to establish a very specific
• 3:53
  working style. We want to override the AI's default
• 3:56
  people -pleasing behavior. You have to give Gemini
• 3:59
  five strict behavioral rules upfront. You tell
• 4:02
  it, number one, answer directly and surely. We
• 4:04
  want pure efficiency. We are cutting out all
• 4:06
  the conversational fluff. Another crucial rule,
• 4:09
  number two, do not default to agreement with
• 4:11
  me. We desperately want independent logic from
• 4:14
  this tool. Yeah, absolutely no yes -men allowed
• 4:16
  in this workflow. It has to challenge you. Rule
• 4:18
  three, no fake critiques, just to sound smart.
• 4:20
  Sometimes the AI invents problems just to seem
• 4:23
  highly analytical. Exactly. Rule four saves us
• 4:26
  from the trap. You tell it. Ask for missing info
• 4:29
  if it changes results. That stops the eager intern
• 4:31
  from just guessing the budget. And finally, rule
• 4:34
  five. Stop when mandatory criteria are met. Do
• 4:38
  not just keep talking and hallucinating new details.
• 4:41
  Two -sex silence. Those five rules completely
• 4:44
  change the interaction dynamic. It forces the
• 4:47
  AI to act like a rigorous partner. It really
• 4:50
  does. Wait, if I have to give it all these strict
• 4:52
  rules, isn't it just faster to write the campaign
• 4:54
  plan myself? It definitely feels that way when
• 4:57
  you first start out, but think about the endless
• 4:59
  revision cycles you normally do. Oh, rewriting
• 5:02
  bad AI traps for three hours straight. Right.
• 5:05
  This upfront alignment eliminates all that massive
• 5:07
  back -end friction. It sets the perfect stage
• 5:10
  for the Gemini work loop. Let's walk through
• 5:12
  that specific framework right now. The work loop
• 5:14
  is a five -step roadmap. for getting things done.
• 5:18
  Step one is simply to understand the actual assignment.
• 5:21
  You make the AI repeat your core goal back to
• 5:23
  you. Step two is to clarify the missing pieces.
• 5:27
  The AI asks you for missing context based on
• 5:29
  your rules. You answer those specific questions
• 5:32
  until the task is completely clear. Step three
• 5:35
  is to explore the possible solutions. The AI
• 5:39
  creates a maximum of three distinct strategic
• 5:41
  directions. It clearly explains the unique strengths
• 5:44
  and trade -offs of each path. It gives you options
• 5:48
  instead of just guessing the best one. Step four
• 5:51
  is the create and revise phase. The AI finally
• 5:54
  builds the deliverable based on your specific
• 5:57
  choice. You review the output and ask for very
• 6:00
  targeted revisions. And step five is the independent
• 6:03
  review phase. You do a final check in a totally
• 6:05
  fresh chat window. It's vital. It's like stacking
• 6:08
  Lego blocks of data. You only snap on the next
• 6:10
  piece when the base is totally solid. That is
• 6:12
  the perfect way to visualize the entire process.
• 6:15
  You are building a stable foundation of locked
• 6:17
  context. A quick access note for our listeners
• 6:20
  today. We are referencing Google AI ProLimits
• 6:22
  in these examples. But this exact core workflow
• 6:25
  applies perfectly to free accounts too. Why do
• 6:28
  we need a totally separate independent review
• 6:30
  step? Because a long chat history will severely
• 6:33
  confuse the AI over time, its attention gets
• 6:36
  diluted across all the rejected ideas. Makes
• 6:39
  sense. A fresh chat prevents old context from
• 6:41
  skewing the final check. Precisely. You want
• 6:44
  a totally clean slate for the final grade. Sponsor,
• 6:47
  host, speaker. We are going to take a brief pause
• 6:50
  here. Sponsor, expert, speaker. All right, and
• 6:54
  we are back to the deep dive. Here he is always
• 6:57
  great in a perfectly clean vacuum, but let's
• 6:59
  see what happens when the pressure is actually
• 7:01
  on. Let's run a real -world scenario. Your manager
• 7:04
  suddenly demands a premium webinar campaign by
• 7:07
  Friday. And you only have some rough audience
• 7:10
  notes to work with. The anxiety is high and the
• 7:12
  deadline is incredibly tight. We're going to
• 7:14
  walk through the practical application of the
• 7:16
  loop. You start by firmly framing and grounding
• 7:19
  the initial request. You do not just blindly
• 7:21
  ask for a campaign plan. You upload your official
• 7:23
  brand guide to the chat window. You provide the
• 7:26
  rough audience notes you actually have. Then
• 7:28
  you force the AI to return two specific lists.
• 7:31
  You ask for confirmed facts versus missing information.
• 7:34
  This completely stops the AI from making assumptions
• 7:37
  right out of the gate. It reads the brand guide
• 7:40
  and your messy notes. It realizes it doesn't
• 7:42
  know the actual webinar budget. or the desired
• 7:45
  speaker lineup. So it asks you for those specific
• 7:48
  details directly. You provide the budget and
• 7:51
  the internal speakers. Now the foundation is
• 7:53
  solid. Next, we move into the explore and choose
• 7:56
  phase. You force the AI to give exactly three
• 7:59
  campaign directions. You demand that each direction
• 8:02
  includes potential risks. You also want to see
• 8:05
  the expected audience reactions for each. This
• 8:07
  is where you actually have to do the hard work.
• 8:09
  The human must choose the direction based on
• 8:12
  real business goals. You cannot outsource your
• 8:15
  strategic judgment to the machine. The AI gives
• 8:17
  you a conservative path, a viral path, and an
• 8:20
  educational path. You look at your quarterly
• 8:22
  goals and pick the educational path. Once you
• 8:24
  pick that path, you create the direction lock.
• 8:27
  This becomes your absolute unbreakable source
• 8:29
  of truth. You literally tell Gemini, do not change
• 8:33
  this direction unless I explicitly approve it.
• 8:35
  It successfully locks in your main goal and target
• 8:38
  audience. It locks in the core promise and exactly
• 8:41
  what topics to avoid. This is mathematically
• 8:44
  important for the language model's attention
• 8:46
  mechanism. You are heavily waiting these instructions.
• 8:49
  With the direction locked, we can finally build
• 8:51
  and check. We build the final campaign brief
• 8:54
  using the Canvas tool. Canvas is a split screen
• 8:57
  workspace to edit long AI drafts. It fundamentally
• 9:00
  changes how you interact with the AI model. It
• 9:03
  moves you from a simple chat bot to a collaborative
• 9:06
  workspace. It keeps your ongoing chat firmly
• 9:08
  on one side. and the living breathing document
• 9:11
  on the other side. You can highlight a single
• 9:14
  problematic paragraph in the document. You ask
• 9:17
  the AI to rewrite just that one tiny section.
• 9:20
  Then you review the total work using incredibly
• 9:22
  strict rules. You label your feedback as pass.
• 9:26
  blocker or optional improvement. You only stop
• 9:29
  the revision process when everything gets a clear
• 9:31
  pass. The brilliance of this blocker rule is
• 9:33
  really fascinating to me. In most organizations,
• 9:36
  feedback is usually incredibly subjective and
• 9:38
  messy. People just say make it pop or I don't
• 9:40
  like this. But a blocker is only valid if it
• 9:43
  violates the direction lock. or if this specific
• 9:46
  claim completely lacks factual evidence. Whoa!
• 9:50
  Imagine scaling this to a billion queries across
• 9:53
  an enterprise. Imagine an entire company giving
• 9:55
  objective, criteria -based feedback to AI. To
• 10:00
  sex silence. You could process immense workloads
• 10:03
  without ever losing the core business goal. It
• 10:06
  forces a level of clarity we rarely demand from
• 10:09
  humans. It forces you to actually define what
• 10:11
  a failure looks like. Is a blocker just anything
• 10:14
  I don't personally like? Not at all. That is
• 10:16
  the beauty of the strict system. Blockers must
• 10:18
  be objective, provable violations of the locked
• 10:21
  brief. Right. Blockers need hard evidence, not
• 10:23
  just a personal preference. Exactly. You are
• 10:25
  enforcing the agreed rules, not just sharing
• 10:28
  your feelings. But there is an obvious bottleneck
• 10:30
  we need to address here. If you run new campaigns
• 10:32
  weekly, typing all this out is exhausting. Typing
• 10:35
  out the five rules and the loop every single
• 10:37
  time is tedious, so we have to talk about how
• 10:39
  we automate the workflow. We need to save this
• 10:41
  rigorous process for future use. You turn your
• 10:44
  proven, tested process into a gem. A gem is essentially
• 10:48
  a packaged, repeatable AI workflow. You create
• 10:52
  a campaign direction architect gem for this specific
• 10:56
  recurring task. It bakes in your grounding logic.
• 10:59
  and your brainstorm format automatically. It
• 11:02
  also includes your strict review rubric and the
• 11:04
  blocker rules. But there is a massive golden
• 11:08
  rule for using these gems. Do not create a gem
• 11:11
  just to save a long prompt. You only create it
• 11:14
  to preserve a process that you know definitively
• 11:17
  works. You must have thoroughly tested the loop
• 11:19
  manually first. If you automate a bad process,
• 11:22
  you just get bad results faster. You also want
• 11:25
  to add new platform features only when absolutely
• 11:27
  needed. You should always start your workflow
• 11:29
  as simply as possible. Start with just the basic
• 11:32
  chat interface. You add files only for highly
• 11:34
  specific necessary source material. You don't
• 11:36
  just dump your entire hard drive into the window.
• 11:39
  You change models only if you need much deeper
• 11:41
  logical reasoning. And you only turn on canvas
• 11:44
  when you have a very long artifact. It is a profound,
• 11:47
  highly effective philosophy of digital minimalism.
• 11:50
  The constant temptation is to use every single
• 11:52
  feature at once. We think more tools will automatically
• 11:55
  give us better results. But that usually just
• 11:58
  creates unnecessary noise in the context window.
• 12:01
  Why shouldn't I just turn on all the extra features
• 12:02
  at once? Adding unnecessary capabilities leads
• 12:06
  directly to AI hallucination or bloated outputs.
• 12:09
  It loses the signal in all that extra noise.
• 12:12
  Keep it simple. Extra tools just mean extra ways
• 12:15
  to get distracted. Spot on. You lose control
• 12:18
  of the loop when you add too much complexity.
• 12:20
  Let's recap the core lesson from today's deep
• 12:22
  dive. You want to use the smallest workflow that
• 12:24
  gives you total control. You must set the AI's
• 12:27
  specific behavior before you begin the real work.
• 12:30
  Fencing in the intern is step one. You lock in
• 12:33
  the context with hard facts and very clear boundaries.
• 12:37
  But you must keep the critical business decisions
• 12:39
  firmly in your own hands. In this specific scenario
• 12:42
  today, we actually had our own context ready.
• 12:45
  We had the brand guide and the rough audience
• 12:47
  notes. But next time, we are jumping into the
• 12:49
  research route. We are going to learn how to
• 12:51
  find, compare, and verify external evidence.
• 12:54
  We will look at what happens when you don't have
• 12:56
  the facts yet. It is about finding the actual
• 12:59
  truth across the wild west of the web. Two sec
• 13:02
  silence. If generative AI is mathematically wired
• 13:06
  to give you the answer that best fits your context,
• 13:08
  is it possible that your biggest obstacle to
• 13:10
  finding the truth is actually your own prompt?
• 13:13
  Out T -Row music.