Latest / AI Fire Daily / #78 Robin: Stop Babysitting Your AI - 5 New Rules for Prompting Claude Opus 5 and Beating Fable 5

Transcript

#78 Robin: Stop Babysitting Your AI - 5 New Rules for Prompting Claude Opus 5 and Beating Fable 5 · AI Fire Daily

• 0:00
  Here are the key conventions to use in an English
• 0:02
  conversational script. One, disfluencies, fillers,
• 0:06
  words like, well, help pace the thought process,
• 0:11
  example. I mean, it's actually a pretty common
• 0:13
  mistake. Two, back -channeling. When one person
• 0:17
  is explaining something, the other person doesn't
• 0:19
  just sit in silence. They offer brief vocal agreements
• 0:22
  or reactions. Use, mm, yeah, right, oh, wow.
• 0:25
  And exactly, example. Host speaker, you have
• 0:27
  to cap the deliverable, expert speaker. Host
• 0:30
  speaker, otherwise it just keeps generating text.
• 0:32
  Three, overlaps and interruptions. Real conversations
• 0:35
  feature people stepping on each other's sentences
• 0:37
  slightly, usually in agreement. Represent this
• 0:39
  with an M dash example, host speaker. It completely
• 0:42
  drains your, expert speaker, cool, your patience,
• 0:45
  right, absolutely. Four, false starts and repetitions.
• 0:48
  People often restart a sentence when formulating
• 0:50
  a complex thought. Example, we had to look at
• 0:53
  the underlying model first. Five, Colloquialisms
• 0:56
  and contractions. Use natural, relaxed phrasing.
• 1:00
  Gonna, kinda, that's it. Example, you're gonna
• 1:04
  wanna skip the heavy verification steps. Six,
• 1:07
  pacing with punctuation. Do not use elixirs.
• 1:10
  Use commas for natural breath pauses, periods
• 1:12
  for definitive stops, and explicit tags like
• 1:15
  beat for purposeful, dramatic pauses. Six, pacing
• 1:18
  with punctuation, natural notes. Claude Opus
• 1:20
  Five is here. Right, but... Your old prompting
• 1:25
  habits. They are about to cost you time. Digital
• 1:28
  patience. And, well, money. Oh, absolutely. It's
• 1:32
  a completely different paradigm now. Welcome
• 1:34
  to the deep dive. We have an essential guide
• 1:36
  for you today. We're mastering Claude Opus 5
• 1:39
  for real knowledge work. Yeah, because treating
• 1:42
  this model like older AI models is just a massive
• 1:45
  mistake. It leads to bloated, slow results. Right.
• 1:49
  We're going to unpack exactly why that happens.
• 1:51
  OK. And we'll reveal the five specific rules
• 1:53
  to fix it. And to prove it works, we have a fun
• 1:56
  test. Oh, yeah. We're going to build a whole
• 1:57
  coffee brand website in three minutes. Just to
• 2:00
  prove it works. Let's start with the foundation.
• 2:02
  Before we even type a prompt, we have to pick
• 2:03
  the right engine. A strong prompt on the wrong
• 2:06
  model yields incredibly weak results. People
• 2:09
  get so paralyzed by the model choices right now.
• 2:13
  You look at the interface, and four options are
• 2:15
  staring back at you. Let's map that out for you.
• 2:17
  First, you have Haiku. That's your casual chat
• 2:19
  model. Then there's Sonnet, which is for daily
• 2:21
  work. That one is the default for the $20 pro
• 2:24
  plan. Right. Then we step up to Opus 5 for complex
• 2:27
  work, planning, and websites. That's the default
• 2:31
  for the $100 max plan. And then there's Fable
• 2:33
  5? Yeah. But only use Fable 5 to test things
• 2:36
  if Opus 5 fails. Really, just leave it on Opus
• 2:39
  5. I want to unpack this effort -level trap.
• 2:42
  Oh, this is a big one. People assume that higher
• 2:45
  effort means a better answer. Right. But effort
• 2:47
  mainly dictates the model's thinking time. For
• 2:50
  normal work, low or medium effort is already
• 2:53
  incredibly strong. It's like using a sledgehammer
• 2:56
  to crack a peanut. That is the perfect analogy.
• 2:59
  High effort is total overkill, unless you're
• 3:01
  tackling a messy research problem or a complex
• 3:04
  code base. Because it just controls how many
• 3:06
  tokens Claude spends, right? Exactly. Can you
• 3:08
  define tokens for us real quick? Sure. They're
• 3:10
  the digital puzzle pieces AI uses to read and
• 3:14
  write text. Okay, so when do you actually upgrade
• 3:16
  a task to Fable 5? You only pull out Fable 5
• 3:19
  when Opus 5 completely misunderstands your prompt's
• 3:22
  core logic. Got it. So, Fable is a fallback for
• 3:25
  logic failures, not a daily driver. Precisely.
• 3:28
  It's just a diagnostic tool. Now that we have
• 3:30
  the right model and effort selected, how do we
• 3:33
  actually talk to it without confusing it? Well,
• 3:36
  older models needed hand -holding. Opus 5 actually
• 3:39
  gets lost if you micromanage it with baby steps.
• 3:42
  Let's bring in our case study. We're looking
• 3:43
  at Northline Coffee? Yeah, they're a small coffee
• 3:46
  bean brand selling online. Rule one is to give
• 3:49
  the full job upfront. instead of asking it to
• 3:52
  extract brand style, then make a plan, then write
• 3:55
  copy. Right. You just declare the final goal
• 3:57
  immediately. You tell it, build a one -page website
• 3:59
  using the provided logo and photos. And save
• 4:02
  it in the project folder. Just give it the final
• 4:04
  destination. But rule two is to define where
• 4:06
  the job ends. Opus 5 is powerful, but it's overly
• 4:10
  eager. Oh, it's a total overachiever. It might
• 4:13
  build a checkout flow you didn't even ask for.
• 4:15
  So you have to give it a box. You say, build
• 4:17
  only the home page. Right. And here's a crucial
• 4:20
  prompt to use. Tell it to make small design choices
• 4:23
  yourself. That is so important. It really is.
• 4:26
  And also say, only ask me a question if the answer
• 4:29
  would change the whole website. I have to admit,
• 4:32
  I still wrestle with micromanaging my prompts
• 4:34
  myself. It's just hard to let go. It's totally
• 4:37
  normal. We're conditioned to do the heavy lifting.
• 4:39
  Why is Opus 5 so prone to over -delivering on
• 4:42
  these tasks, though? It comes down to its training
• 4:45
  data. It's essentially built to be the ultimate
• 4:48
  people -pleaser. It predicts your future needs.
• 4:51
  Right. It over -delivers because its training
• 4:54
  heavily rewards being aggressively helpful. Exactly.
• 4:57
  So we've given it the job and the boundaries,
• 5:00
  but what happens when it drops a giant wall of
• 5:03
  text on you? That is the worst part. Mid -roll
• 5:06
  sponsor read. We're back. Let's talk about setting
• 5:09
  guardrails. We have to distinctly control the
• 5:12
  AI's talking versus its actual doing. This is
• 5:15
  a massive pain point. Opus 5 loves to explain
• 5:18
  every single choice it made. In a massive wall
• 5:20
  of text that no one actually reads. Nobody reads
• 5:22
  it. So rule three is limit the text response.
• 5:25
  You just tell it, give me no more than three
• 5:27
  short bullets explaining what you made. It saves
• 5:29
  so much cognitive load. But rule four is limiting
• 5:32
  the work product itself. That's different from
• 5:35
  the chat response. Totally different mechanics.
• 5:37
  You must cap. the actual deliverable. For Northline
• 5:40
  coffee, you limit the site to exactly six parts.
• 5:43
  Right. The hero section, three featured coffees,
• 5:46
  why we're different, reviews, FAQ, and a buy
• 5:49
  button. You cap the copy too. Headlines under
• 5:52
  eight words, sections under 70 words. Or you
• 5:55
  can just say, Make it readable in under 90 seconds.
• 5:58
  That's a great proxy. Do strict word counts end
• 6:01
  up stifling the AI's creativity, though? The
• 6:04
  exact opposite, actually. Unlimited space leads
• 6:07
  to lazy, generic filler. Strict constraints actually
• 6:10
  force the AI to choose the most impactful words.
• 6:13
  Exactly. It forces prioritization. We've boxed
• 6:16
  in the size, but how do we make sure the quality
• 6:18
  is there without bogging down the system? That
• 6:20
  brings us to rule five. Don't tell it to double
• 6:23
  -check everything twice. Ah, the old safety blanket.
• 6:27
  Yeah, older prompt advice always said, check
• 6:28
  every file twice. Find every mistake. With Opus
• 6:32
  5, doing that just wastes tokens and time. Because
• 6:35
  it auto -adjusts while working. It's already
• 6:37
  verifying its own logic. So instead, you just
• 6:39
  vividly describe what good work feels like. Right.
• 6:42
  You say, make it clean, premium, and easy to
• 6:45
  scan. Let's talk about the final test. We pasted
• 6:48
  this full prompt and the assets into Claw Design.
• 6:51
  We set it to Opus 5 on medium effort. And the
• 6:54
  result was pretty stunning. Whoa. Beat. Imagine
• 6:58
  building a fully scoped on -brand website in
• 7:01
  under three minutes with zero follow -up questions.
• 7:04
  That's wild. It matched the exact six sections.
• 7:08
  They respected all the word limits perfectly.
• 7:10
  And it output exactly three bullet points in
• 7:12
  the chat. It's staggering efficiency. Zero time
• 7:15
  wasted guessing your intentions. Does relying
• 7:18
  on AI this much remove the human element of design?
• 7:22
  Not at all. It just removes the repetitive formatting
• 7:24
  work. You become the core strategist dictating
• 7:26
  the vision. It eliminates repetitive formatting
• 7:28
  but elevates you to the core strategic director.
• 7:31
  Precisely. Let's recap the big idea here. Two
• 7:34
  -sec silence. The goal of Opus 5 is to let it
• 7:37
  use its immense intelligence without letting
• 7:39
  it overuse that intelligence. Yes. Set the box,
• 7:42
  then let the AI play inside it. Give it the full
• 7:45
  job. Set the finish line, restrict the size,
• 7:47
  and let it work. It leaves us with a rather profound
• 7:50
  question, though. What's that? If AI models like
• 7:52
  Opus 5 no longer need us to hold their hands
• 7:56
  through microsteps, and they can self -correct
• 7:58
  their own work, What does our role become? Right,
• 8:01
  and how does our value shift? Exactly. Are we
• 8:04
  moving from being the creators of knowledge work
• 8:06
  to acting purely as the directors of digital
• 8:10
  talent? It's a huge shift to think about. Something
• 8:12
  to mull over the next time you open a blank prompt
• 8:15
  window. Thank you for taking this deep dive with
• 8:17
  us.