Latest / AI Fire Daily / #88 Robin: Make Claude Speak Human - The ELI5 Output Style

Transcript

#88 Robin: Make Claude Speak Human - The ELI5 Output Style · AI Fire Daily

• 0:00
  You ask an AI a simple question and it hands
• 0:03
  you a textbook. Why is the smartest technology
• 0:06
  so hard to talk to? Yeah, that is the great irony
• 0:09
  of this whole tech cycle. We've managed to build
• 0:12
  these unbelievably powerful reasoning engines,
• 0:15
  yet we struggle to just have a normal frictionless
• 0:18
  conversation with them. Right. Well, welcome
• 0:21
  to this deep dive. I'm really glad you were here
• 0:23
  with us. Me too. We are looking at a brilliant
• 0:25
  guide today. It's all about how to fix Claude
• 0:29
  AI's communication style inside Claude code.
• 0:32
  Yeah, because raw intelligence is... practically
• 0:35
  useless if you can't extract the insights cleanly.
• 0:38
  You end up, you know, working for the AI instead
• 0:40
  of it working for you. Exactly. So here is the
• 0:43
  roadmap for our journey today. First, we are
• 0:45
  going to separate the two core problems holding
• 0:48
  back daily AI use. Which are the heavy jargon
• 0:51
  and the massive length. Right. Then we will build
• 0:53
  a permanent foundation for clear language. And
• 0:56
  finally, we will introduce some on -demand shortcuts
• 0:58
  to tame those huge walls of text. It's a highly
• 1:02
  practical framework. We are moving away from
• 1:04
  just complaining about the model. We are stepping
• 1:06
  into actual structural control over how it behaves.
• 1:09
  So let's unpack that core misunderstanding first.
• 1:12
  I hear this complaint all the time from people
• 1:14
  using Opus 5. They say, Claude just talks way
• 1:17
  too much. Yeah, everyone says that. But the guide
• 1:19
  points out something fascinating. That complaint
• 1:22
  is actually two entirely different problems.
• 1:24
  They are just masquerading as one. They get lumped
• 1:27
  together because the symptom feels exactly the
• 1:29
  same. You just feel overwhelmed. Right. But if
• 1:32
  you try to treat them as a single bug, you actually
• 1:34
  end up breaking the AI's ability to think. You
• 1:37
  really have to isolate the jargon from the link.
• 1:40
  Let's look at the jargon problem first. Opus
• 1:42
  5 is a genius level model, but it constantly
• 1:46
  assumes you have an expert level vocabulary.
• 1:49
  Right. In whatever subject you bring up. It wants
• 1:52
  to demonstrate its technical competence. So it
• 1:55
  defaults to this incredibly dense academic language.
• 1:59
  It doesn't know how to casually chat by default.
• 2:01
  The source gives this perfect, relatable example.
• 2:06
  Say you are looking at a business report. You
• 2:09
  ask Claude a really simple question. Our CRM
• 2:12
  open rate is 32%. Are open rates basically fake
• 2:16
  now? That is a totally normal question. You're
• 2:18
  just trying to get a quick gut check. Exactly.
• 2:21
  You want a yes, a no, and maybe two sentences
• 2:23
  of context. But Claude doesn't give you that.
• 2:25
  Instead, it hits you with heavy technical jargon.
• 2:29
  It says things like, an Apple Mail recipient
• 2:31
  with MPP on registers as an open. Right. And
• 2:34
  then it starts breaking down geodevice data derived
• 2:37
  from opens. Which is technically accurate, sure.
• 2:40
  But it's completely misplaced. You asked for
• 2:42
  quick business advice. Yeah, and it gave you
• 2:44
  a software engineering anatomy lesson. It is
• 2:46
  like asking someone for the time, and they explain
• 2:49
  how to build a Swiss watch. That is the perfect
• 2:51
  way to frame it. It's exhausting. It literally
• 2:53
  gives you homework. You have to sit there and
• 2:55
  decode the AI's answer before you can actually
• 2:57
  do anything with it. And that leads right into
• 2:59
  the second issue, the wall of text problem. Getting
• 3:02
  a five paragraph essay for a simple question
• 3:04
  has two major costs. Time is the first one. Yeah.
• 3:08
  You lose momentum in your workflow because you
• 3:10
  are reading a mini dissertation. And the second
• 3:12
  cost is literal money. It wastes your tokens.
• 3:16
  Let's define that quickly. Tokens are small pieces
• 3:18
  of words and AI reads and writes. Right. And
• 3:21
  every time Claude generates a massive wall of
• 3:23
  text, you are paying for those tokens through
• 3:27
  the API. So a rambling AI is literally draining
• 3:30
  your account balance. While slowing you down.
• 3:32
  Exactly. Let me ask you this. Why do incredibly
• 3:36
  smart models like Opus 5 default to this dense,
• 3:40
  heavy jargon? Well, it stems from trying to be
• 3:43
  technically precise. It assumes you already know
• 3:46
  the field. So it leads with complex mechanisms
• 3:49
  rather than the simple truth. So it prioritizes
• 3:52
  technical precision over human readability. Got
• 3:54
  it. Yeah, it fundamentally believes that precision
• 3:56
  equals helpfulness. It doesn't realize that for
• 3:59
  a human, clarity equals helpfulness. Which brings
• 4:02
  us to the first major fix in the guide, the custom
• 4:05
  ELI -5 output style. This is where we rewire
• 4:08
  the foundation. Because if the AI naturally speaks
• 4:11
  in dense code... They're telling you to be shorter
• 4:13
  doesn't help. Exactly, you just get shorter dense
• 4:15
  code. We have to fix the foundational voice first.
• 4:18
  We do that by defining an output style. An output
• 4:21
  style is the permanent writing voice the AI uses
• 4:23
  when replying. Most people don't even know this
• 4:25
  exists. They just use whatever default personality
• 4:28
  the engineers shipped it with. And that default
• 4:30
  style is formal. It leans super heavily toward
• 4:33
  technical completeness. The guide suggests intercepting
• 4:36
  that behavior. Right. By creating a custom style.
• 4:40
  Specifically, an ELI -5 style. Explain it first
• 4:44
  roll. Like I'm five. I love this approach. But
• 4:47
  I have to admit, I still wrestle with Cromp Drift
• 4:50
  myself. My instructions just fade away. It is
• 4:52
  incredibly frustrating, yeah. You spent all this
• 4:54
  time crafting perfect rules. And an hour later,
• 4:57
  the AI acts like it never met you. Exactly. Prompt
• 5:01
  drift is when an AI forgets initial rules over
• 5:03
  a long chat. I used to put all my communication
• 5:05
  rules into a file called Claude .md. Right, for
• 5:08
  project rules. But it just never stuck for very
• 5:10
  long. The guide actually breaks down why that
• 5:12
  happens. A Claude .md file is excellent for project
• 5:16
  -level rules, like repository notes. But it gets
• 5:19
  loaded near the beginning of your session, and
• 5:21
  as you chat, you are stacking more text on top
• 5:24
  of it. It is like your desk getting covered in
• 5:26
  new papers. Eventually those foundational rules
• 5:29
  get buried. The model just drifts away from them.
• 5:32
  But an output style operates differently under
• 5:35
  the hood. It sits much closer to the core operational
• 5:38
  instructions. It's a persistent layer that locks
• 5:41
  in the behavior. Constantly reminding Claude
• 5:43
  how to speak on every single turn. Let's get
• 5:46
  into the mechanics of this ELI -5 style. The
• 5:49
  guide gives a very specific structural framework
• 5:52
  for it. You don't just say talk simply. No, you
• 5:54
  give it a logical progression. First... lead
• 5:57
  with the key answer. Right. Second, explain it
• 5:59
  in plain language. Third, give the next action.
• 6:02
  Yeah. And fourth, only then do you add the technical
• 6:05
  details. You are forcing it to put the human
• 6:07
  takeaway at the very top. So if we go back to
• 6:09
  our CRM example, with this ELI -5 style active,
• 6:14
  QUAT's response changes entirely. It starts with,
• 6:16
  that 32 % is not all real people. Boom. Instant
• 6:20
  clarity. the cognitive load just drops to zero.
• 6:23
  Then it adds the context gently. It says some
• 6:25
  email apps trigger opens automatically. You don't
• 6:28
  have to translate geodevice data anymore. The
• 6:31
  AI is still doing the complex analysis. It is
• 6:33
  just delivering it in a human compatible format.
• 6:36
  There's an optional upgrade to this rule that
• 6:38
  I found absolutely fascinating. You can add a
• 6:40
  standard called ASD -STE100 to the prompt. Simplified
• 6:44
  technical English. It's a brilliant hack. It
• 6:47
  originates from the aerospace industry, right?
• 6:48
  Yeah, they created it so mechanics wouldn't misinterpret
• 6:51
  maintenance manuals. You really don't want a
• 6:53
  technician getting confused by vague verbs on
• 6:55
  a jet engine. Definitely not. So STE forces the
• 6:59
  use of a very constrained, approved vocabulary.
• 7:02
  And applying this to AI acts as a rigid filter.
• 7:06
  It stops the agent from using ambiguous jargon.
• 7:08
  It pairs beautifully with ELI -5. ELI -5 gives
• 7:12
  you the structure and tone. STE gives you the
• 7:14
  strict vocabulary constraints. Why does using
• 7:16
  the exact phrase ELI -5 work so effectively on
• 7:19
  an AI? Well, because AI models have digested
• 7:22
  vast amounts of Internet data. And in that data,
• 7:25
  ELI -5 is a proven pattern for explaining hard
• 7:27
  topics simply. It triggers a deep pattern of
• 7:30
  simple teaching in its training data. Exactly.
• 7:32
  You are leveraging the collective teaching knowledge
• 7:35
  of the Internet. We're going to take a quick
• 7:37
  break. When we come back, we will tackle the
• 7:39
  second half of the equation. How to handle the
• 7:41
  massive wall of text. Mid -roll sponsor, Red
• 7:45
  Marker. Oh. OK, we're back. We've established
• 7:48
  a solid foundation. The AI is now speaking in
• 7:51
  clear, plain English thanks to our ELI5 output
• 7:54
  style. The jargon is gone. It finally sounds
• 7:57
  like a helpful human. But now we face the length
• 7:59
  issue. We still have the wall of text problem.
• 8:02
  What happens when we actually need the AI to
• 8:04
  be brief? The immediate instinct for most people
• 8:07
  is just to update their system prompt. They type
• 8:09
  in, always be concise, and think they fixed it.
• 8:12
  Right. But the guide warns against this heavily.
• 8:14
  Let me ask you this. Why is it so dangerous to
• 8:17
  just put always be concise in the main system
• 8:19
  prompt? Uh, well, if you force an AI to be permanently
• 8:22
  brief, you actually cripple its intelligence.
• 8:25
  How so? Large language models think by generating
• 8:28
  tokens. They use step -by -step reasoning. So
• 8:30
  if you cap their output... You are literally
• 8:33
  cutting off their cognitive process. Exactly.
• 8:35
  They can't think through complex debugging or
• 8:38
  deep research anymore. You keep the deep reasoning
• 8:40
  available, but hide it until you ask for it.
• 8:43
  That's the secret. You need that deep reasoning
• 8:46
  to happen. You just don't want to read a novel
• 8:48
  every time. So the solution here is skills. These
• 8:51
  are reusable shortcut commands that you call
• 8:54
  only when you need them. Think of the output
• 8:56
  style as setting your car suspension to comfort
• 8:59
  for the whole road trip. OK. The skills are like
• 9:02
  having a turbo button you press just when you
• 9:04
  need to pass a truck. They are temporary overrides.
• 9:08
  Let's look at the first skill. The naming convention
• 9:10
  here made me laugh. It is called slash bro, just
• 9:14
  slash B -R -O. It is highly intuitive. You know
• 9:17
  exactly what that button does. The command is
• 9:19
  incredibly simple. It tells the AI to restate
• 9:22
  its last message in plain human language. You
• 9:25
  deploy this when Claude inevitably slips back
• 9:28
  into dense technical speech, because no matter
• 9:30
  how good your prompt is, an AI will occasionally
• 9:33
  revert. So instead of getting frustrated, you
• 9:35
  just type slash bro. And the mechanism is brilliant.
• 9:39
  The AI looks at its own previous dense output.
• 9:43
  It runs that text through a simplification filter
• 9:46
  and hands you the translated version. It's an
• 9:48
  emergency break for technical density. Now, the
• 9:52
  second skill is called, wait, what? This one
• 9:55
  is vital for complex workflows. It is focused
• 9:58
  on context, not just vocabulary. You use this
• 10:01
  when you completely lose the thread of the conversation.
• 10:04
  Yeah, maybe Claude jumped into writing a Python
• 10:06
  script way too fast. Or it made an assumption
• 10:08
  about your database that you didn't catch. You're
• 10:11
  staring at the screen feeling totally lost. You
• 10:12
  just thought, wait, what? The skill forces Claude
• 10:15
  to stop and repitch the idea. It explains where
• 10:17
  we currently are and what the next step is. Now,
• 10:20
  there's a crucial caveat from the source material
• 10:23
  here. You should never blindly copy other people's
• 10:27
  context skills. Right. People find a cool script
• 10:29
  on GitHub and just paste it in, but that script
• 10:31
  might reference specific team documents that
• 10:33
  you don't actually have. If you copy that, Claude
• 10:36
  will start hallucinating. You'll go looking for
• 10:38
  a PDF that simply doesn't exist on your hard
• 10:41
  drive. So you have to ask Claude to adapt the
• 10:43
  wait what command specifically for your own workspace.
• 10:47
  That grounds the explanation in your reality.
• 10:50
  Now, the third skill is definitely my favorite.
• 10:53
  It's the slash quick command slash quick n. It
• 10:56
  is the ultimate compression tool. It compresses
• 10:59
  the last answer into exactly n. key points, and
• 11:02
  it explicitly ranks them by importance. So if
• 11:05
  you're looking at a massive wall of text, you
• 11:08
  just type slash quick three. The AI processes
• 11:11
  the whole essay and gives you the top three takeaways.
• 11:14
  Or if you type slash quick one, you bypass everything.
• 11:16
  You just get the single most important action
• 11:18
  item. You let Claude write the full essay in
• 11:21
  the background. You let it do all that vital
• 11:23
  deep reasoning. And then you just extract the
• 11:25
  juice. Right. You don't stop the thinking process.
• 11:27
  You just aggressively filter the final output.
• 11:30
  Whoa. Two secs silence. Imagine holding this
• 11:32
  exact control layer over every future AI model,
• 11:36
  bending it entirely to your workflow. It is a
• 11:39
  profound shift in power dynamics. You dictate
• 11:42
  the terms of engagement now. It makes you the
• 11:44
  director of the intelligence, not just a passive
• 11:46
  passenger. And as new models drop, you just lay
• 11:49
  this control framework over them. You tame them
• 11:51
  instantly. Let's zoom out for a moment. This
• 11:53
  brings us to the... Big idea of this entire deep
• 11:56
  dive. The underlying philosophy of human AI interaction.
• 12:00
  The guide makes a really salient point here.
• 12:03
  High benchmark scores do not equal good. daily
• 12:07
  usability. It is a massive blind spot in the
• 12:10
  tech industry right now. Everyone is obsessing
• 12:12
  over benchmarks. We look at math scores and coding
• 12:15
  evaluations. Opus 5 crushes those objective tests.
• 12:19
  But raw capability is meaningless if the communication
• 12:21
  style causes friction in your actual workday.
• 12:24
  Friction is the enemy of adoption. If an AI gives
• 12:27
  you a mathematically perfect answer, but you
• 12:30
  need 10 minutes to decode it, the tool has failed
• 12:33
  you. It becomes a net loss of productivity. You
• 12:35
  honestly would have been faster just doing the
• 12:37
  research yourself. The human cost of extraction.
• 12:39
  That is the real metric we should be optimizing
• 12:41
  for. Absolutely. So the ultimate formula for
• 12:44
  daily AI work is this. Clarity belongs in your
• 12:47
  default output style. That's your permanent ELI
• 12:50
  -5 foundation. And brevity belongs in your on
• 12:52
  -demand shortcuts. Those are your transient skills.
• 12:56
  Your slash quick and slash bro commands. You
• 12:58
  fundamentally separate how the AI speaks from
• 13:01
  how much the AI speaks. Applying that separation
• 13:04
  solves almost every frustration people have with
• 13:06
  these agents. It requires a complete mental shift.
• 13:10
  You are engineering the communication layer,
• 13:12
  not just prompting for a task. And as better
• 13:14
  models keep coming, they will always have bizarre
• 13:17
  quirks. But if you master this control layer,
• 13:19
  you are insulated from those quirks. So here's
• 13:22
  our call to action for you today. Take 15 minutes.
• 13:25
  Open up Claude Code. and set up an ELI -5 style
• 13:28
  for your own workspace. It is a tiny investment
• 13:31
  of time for a massive return. Test it out. Notice
• 13:34
  how it dramatically changes your daily reading
• 13:36
  time. You'll feel the cognitive load lift immediately.
• 13:39
  It is remarkably refreshing to just get a straight
• 13:42
  answer. And remember to apply the framework.
• 13:44
  Always separate clarity from brevity. Let the
• 13:46
  AI think deeply, but make it speak plainly. I
• 13:50
  want to leave you with a final thought today.
• 13:52
  Something to ponder as these complex systems
• 13:54
  continue to evolve. Right now, we are spending
• 13:57
  all this effort to force AI to speak plainly
• 14:00
  to us. We are imposing strict human limitations
• 14:03
  on its natural output. Exactly. But what happens
• 14:06
  in a few years? What happens when these autonomous
• 14:09
  AI agents are communicating primarily directly
• 14:12
  with each other? When humans are completely removed
• 14:14
  from the loop? Will they develop their own hyperdense
• 14:17
  jargon? A language so compressed we can't even
• 14:19
  monitor it? Or will they actually need their
• 14:21
  own internal ELI -5 protocols just to keep their
• 14:25
  own operations running smoothly? That is a terrifying
• 14:27
  and utterly fascinating question. It really is.
• 14:30
  Thank you for joining us on this deep dive. Go
• 14:32
  set up those ELI -5 rules and we will catch you
• 14:35
  next time.