Latest / AI Fire Daily / #108 Robin: Editing Videos with Code, GPT-6 Astra, and Plain English

Transcript

#108 Robin: Editing Videos with Code, GPT-6 Astra, and Plain English · AI Fire Daily

• 0:00
  Imagine staring at a massive 14 minute video
• 0:03
  timeline. You know, you are painfully hunting
• 0:06
  for every dead space. Yeah, dragging your face
• 0:08
  cam frame by frame. Right. Beat. Welcome to the
• 0:12
  deep dive. Today, we are exploring a workflow
• 0:15
  that kills that manual grind entirely. And it
• 0:18
  really is a necessary evolution, honestly. We
• 0:20
  are turning raw footage into a polished edit
• 0:22
  using only plain English commands. I mean, traditional
• 0:25
  editing is just this agonizing manual reality.
• 0:29
  You're constantly cutting out stutters. And syncing
• 0:32
  captions entirely by hand. Exactly. You hunt
• 0:34
  endlessly for relevant B -roll, and then you
• 0:37
  spend hours building motion graphics. It is just
• 0:40
  purely exhausting. But that reality is shifting
• 0:42
  fast, which is why we are here. We are unpacking
• 0:45
  a guide today that covers a very specific AI
• 0:47
  video editing stack. Yeah, we are looking at
• 0:49
  codecs, GBT -6, Astra, and hyperframes. Right.
• 0:53
  Our mission is to deeply understand how this
• 0:55
  actually works. First, we will look at what this
• 0:58
  AI can output today. Which is pretty mind blowing.
• 1:00
  It is. Second, we will break down the five -step
• 1:02
  core workflow. Third, we will set up the local
• 1:05
  workspace. get in your hands duty. Exactly. And
• 1:08
  finally, we will do a hands -on walkthrough of
• 1:11
  a real edit. It is a phenomenal roadmap. I am
• 1:15
  really excited to dig into the mechanics of this.
• 1:18
  So let's start with what AI video editing can
• 1:20
  actually do today. Before we look under the hood
• 1:22
  at the software, we have to ask the obvious question.
• 1:25
  Is the final product actually worth watching?
• 1:27
  Right. Or is it just another gimmicky tech demo?
• 1:30
  That is the ultimate test, right? And well, the
• 1:33
  results are actually stunning. Let's look at
• 1:34
  long form YouTube editing first. Yeah. The guide
• 1:37
  highlighted a test using a 14 minute raw recording.
• 1:40
  Right. It was a standard mix of a face cam and
• 1:43
  a screen recording tutorial. Yeah. And the AI
• 1:46
  took that 14 minutes down to nine minutes. Wow.
• 1:48
  So it automatically stripped out all the stutter.
• 1:50
  Yeah. It removed all the dead space entirely.
• 1:52
  Saving five minutes of awkward pauses is just
• 1:55
  huge. It really is. But the intelligent That
• 1:57
  layout logic is what really blew me away here.
• 2:00
  Yeah, that layout logic is genuinely fascinating.
• 2:02
  Let's say the screen recording is totally static.
• 2:06
  Like, you are explaining a concept, but nothing
• 2:08
  is happening on screen. Exactly. The AI recognizes
• 2:11
  that empty space. It realizes the screen is not
• 2:15
  providing value right then. Right, so it expands
• 2:18
  your facecam to full screen automatically. Which
• 2:20
  is incredibly smart. Then, when the tutorial
• 2:23
  content matters again, it shrinks the facecam
• 2:25
  back down. It seamlessly moves the focus back
• 2:29
  to the screen recording. That behavior comes
• 2:31
  from incredibly clear instructions, but it also
• 2:34
  comes from what the guide calls reusable skills.
• 2:37
  Oh, yeah. That is massive. You build a preferred
• 2:40
  aesthetic once and the AI remembers it forever.
• 2:43
  Right. And that changes everything for independent
• 2:44
  creators. You just you don't have to reinvent
• 2:47
  the wheel every time. But it also handles short
• 2:49
  form reels beautifully, right? Oh, absolutely.
• 2:51
  This is where the underlying workflow gets much
• 2:53
  more aggressive. the pacing changes completely
• 2:56
  for reels. Right. It starts changing scenes every
• 2:58
  one to two seconds. Yeah. And it dynamically
• 3:01
  mixes full -screen speaker shots with b -roll.
• 3:03
  It adds layered 3D motion graphics on top of
• 3:05
  that. And the subtitles are perfectly synced
• 3:08
  to the exact spoken words. Exactly. It generates
• 3:11
  animated text on the fly. The thematic visual
• 3:14
  mapping is what really caught my eye there. Like,
• 3:16
  one reel used bright orange accents for a topic
• 3:20
  about Claude. Yeah. And then another reel used
• 3:22
  green accents for a topic about Nvidia. Right.
• 3:25
  It is actively matching the brand colors of the
• 3:28
  subject matter. It keeps the visuals moving continuously.
• 3:31
  It never leaves static elements on screen for
• 3:33
  too long. Which retains viewer attention. Precisely.
• 3:36
  The AI essentially plans the visual pacing around
• 3:39
  the topic itself. But it is not just... fast
• 3:42
  TikTok cuts, which can get exhausting, the dynamic
• 3:45
  range is quite impressive. Yeah. Take a look
• 3:48
  at the slower call -to -action videos. Those
• 3:50
  require a totally different much softer touch.
• 3:53
  Right. For a calm call to action, the AI used
• 3:55
  much slower pacing. It used low ambient background
• 3:58
  music. It added gently cropped floating screenshots
• 4:01
  of courses and resources. So the key takeaway
• 4:04
  is really crucial here for you listening. The
• 4:06
  AI adapts to the actual purpose of the scene.
• 4:09
  Exactly. It does not force the exact same chaotic
• 4:12
  editing style everywhere. The quality really
• 4:14
  comes from those reusable skills. Right, you
• 4:17
  lock in your specific visual preferences and
• 4:19
  pacing rules. And then you just deploy them instantly
• 4:21
  on the next video. This adaptive pacing is incredibly
• 4:24
  interesting to me. Beat. Does this adaptive pacing
• 4:27
  mean the AI actually understands the context
• 4:30
  of the speech, or is it just reacting to audio
• 4:32
  levels? Oh, it definitely understands the semantic
• 4:34
  context. Yeah. GPT -6 Astra is reading the transcript
• 4:38
  deeply. It knows when you are explaining a complex,
• 4:43
  dense topic versus when you are making a quick
• 4:45
  joke. It paces the visuals based on that actual
• 4:48
  underlying meaning. So it actually maps the visual
• 4:51
  rhythm directly to the semantic meaning. Got
• 4:54
  it. This is where we really see how the magic
• 4:55
  happens. Yeah. To achieve that level of dynamic,
• 4:59
  context -aware editing, we need strict structure.
• 5:02
  Absolutely. The AI has to follow a very specific
• 5:05
  logical sequence. If it doesn't, everything falls
• 5:07
  apart. That sequence is transcribe, cut. Plan,
• 5:11
  generate, and verify. Right. Step one is always
• 5:14
  transcription. The AI converts the spoken audio
• 5:16
  into text. But crucially, it creates precise,
• 5:20
  word -level timestamps. That timing matters immensely,
• 5:23
  doesn't it? Oh, entirely. Subtitles and visual
• 5:26
  changes must match the exact spoken word. If
• 5:28
  they drift by even half a second, the video feels
• 5:31
  completely unwatchable. Yeah. You have two main
• 5:34
  options for this transcription step. Whisper.
• 5:37
  or Eleven Labs. Whisper is great because it is
• 5:40
  a completely free local option. It runs on your
• 5:42
  own hardware. Right, and Eleven Labs requires
• 5:45
  a paid setup, but it is substantially faster.
• 5:47
  Exactly. After transcribing, you move to step
• 5:50
  two. You cut. You must remove the mistakes, the
• 5:53
  dead space, and the long pauses right now. Yeah,
• 5:55
  this is a critical sequencing rule that you cannot
• 5:57
  ignore. You have to cut the dead space before
• 6:00
  planning any visual elements. I want to pause
• 6:02
  on that. Why is that specific order so strictly
• 6:05
  required? Because if you cut the footage after
• 6:08
  creating the visual plan, all your timestamps
• 6:10
  will completely shift. Ah, I see. If you remove
• 6:13
  three seconds of audio, every animation after
• 6:15
  that point is now three seconds late. Wow. The
• 6:18
  animations will suddenly lose sync with the speech.
• 6:21
  It creates absolute chaos in the final edit.
• 6:23
  That makes perfect sense. You have to finalize
• 6:25
  the timeline length first, then step three is
• 6:28
  planning the beats. Yeah, a beat is essentially
• 6:31
  one distinct visual moment or scene. GPT -6 Astra
• 6:35
  reads that freshly cleaned transcript. It actively
• 6:39
  decides what should appear on the screen at any
• 6:41
  given moment. Exactly. It plans exactly when
• 6:43
  sound effects and visual changes should happen
• 6:45
  based on the words. Then we hit step four. Generate.
• 6:48
  This is where Hyperframe steps in. It takes that
• 6:51
  text -based plan and turns it into animated HTML.
• 6:54
  Code that structures web pages and digital visual
• 6:56
  elements. Exactly. It uses web technologies to
• 6:59
  render the video instead of a traditional timeline.
• 7:02
  Whoa. Two secs silence. Building the visual layer
• 7:05
  entirely from code instead of a timeline? That
• 7:08
  is just wild. It really is a paradigm shift.
• 7:10
  I mean, it is not manually rendering massive
• 7:13
  pixel files. It is generating structural programmatic
• 7:16
  code. Right, it creates the motion graphics and
• 7:19
  3D elements natively in code. This makes it incredibly
• 7:22
  lightweight and infinitely adjustable. Then we
• 7:25
  reach step five, the verify loop. Yeah, the agent
• 7:28
  actually takes screenshots of its own generated
• 7:30
  work. It checks those images against the transcript
• 7:33
  again. It looks for elements clipping outside
• 7:35
  the frame. It checks for bad timing or incorrect
• 7:38
  visual sync. And if something looks wrong in
• 7:40
  the screenshot, it updates the underlying code
• 7:43
  and checks again. Exactly. It is a self -correcting
• 7:45
  system. With the verification loop, how many
• 7:48
  passes does it typically take before the agent
• 7:50
  stops tinkering? It varies depending on complexity.
• 7:54
  but usually just a few passes. It keeps adjusting
• 7:57
  the HTML and checking the new screenshots. Until
• 8:00
  it gets it right. Yeah, it constantly compares
• 8:02
  the visual output to your original text instructions
• 8:04
  until it aligns perfectly. It loops until the
• 8:07
  pixels match the prompt's exact constraints.
• 8:10
  Makes perfect sense. It really does. Now that
• 8:12
  we understand the theory of the loop, how do
• 8:14
  we actually get this running? We need to install
• 8:16
  these tools on our own desk. It sounds intimidating,
• 8:19
  but the setup is actually quite logical. You
• 8:22
  only need three main pieces to start, Codex,
• 8:25
  Hyperframes, and 11 Labs. Right. Step one is
• 8:28
  creating a local project folder on your machine.
• 8:31
  You literally just make a new folder on your
• 8:33
  desktop. You might name it Hyperframes demo or
• 8:36
  something simple. Yeah. Then you open the Codex
• 8:39
  desktop app. You point Codex directly to that
• 8:41
  newly created folder. That folder basically becomes
• 8:44
  the brain of your project. It holds your raw
• 8:47
  video files and your final renders. Exactly.
• 8:49
  Step two is installing Hyperframes into that
• 8:52
  brain. You copy the Hyperframes GitHub repository
• 8:55
  URL. You paste it right into Codex. Right. You
• 8:58
  simply ask Codex to pull the repo into the project
• 9:01
  folder. Codex acts like an assistant here. It
• 9:04
  installs all the required dependencies automatically?
• 9:06
  Yeah, it builds the project files for you. And
• 9:09
  crucially, it gives you a local host preview
• 9:11
  link. That localhost preview acts as your monitor.
• 9:14
  It is where you inspect and test the video while
• 9:17
  you work. You watch the code compile individuals
• 9:19
  in real time. It is really cool. Step three is
• 9:22
  connecting 11 Labs for the transcription engine.
• 9:24
  You go into your 11 Labs account and create a
• 9:27
  new API key, a secret password allowing two pieces
• 9:31
  of software to talk safely. You should restrict
• 9:33
  that key to speech to text only at first, just
• 9:36
  for security. Treat it exactly like a sensitive
• 9:39
  password. Never paste it into public prompts
• 9:41
  or share it online. Exactly. Step four is storing
• 9:45
  that API key safely in your project. You ask
• 9:48
  Codex to create a .env file. You paste the key
• 9:51
  into that hidden file. This keeps it secure but
• 9:54
  accessible to the local agent. Finally, we reach
• 9:57
  step five. You add your project rules to a file
• 10:00
  called agents .md. Yeah, this file acts as the
• 10:04
  master instruction manual for your specific project.
• 10:06
  I have to admit something here. Beat in. I still
• 10:09
  wrestle with prompt drift myself. Oh, we all
• 10:11
  do. It is incredibly hard to keep AI outputs
• 10:15
  consistent over time. You give an instruction
• 10:17
  on Monday, and by Thursday, the AI has completely
• 10:21
  wandered off and changed its style. Right. And
• 10:23
  that is exactly why the agents .md file exists.
• 10:26
  It permanently anchors the AI. You tell it to
• 10:28
  always use 11 labs for transcription inside that
• 10:30
  file. You store your preferred editing, pacing,
• 10:32
  and visual style right there. You save your folder
• 10:35
  structures and your specific visual preferences.
• 10:37
  The more rules you save in that document, the
• 10:39
  less you have to repeat yourself later. Exactly.
• 10:41
  It remembers for you. If agents .md holds all
• 10:44
  the rules, does it get bloated or is it like
• 10:47
  stacking Lego blocks of data? It is very modular,
• 10:51
  much like Lego blocks. You just add discrete
• 10:53
  rules one by one as you discover them. Oh, that
• 10:55
  makes sense. You lock in your font style block,
• 10:57
  then snap on your pacing block. It stays incredibly
• 11:00
  clean and organized if you manage it well. Right,
• 11:02
  you just build it block by block as your style
• 11:04
  evolves. We will take a quick break here. Sponsor.
• 11:08
  And we are back! We have the tools installed
• 11:11
  and the rules set up. Let's see how this stack
• 11:13
  handles a messy real -world editing scenario.
• 11:17
  Yeah, the guide walks through a raw 65 second
• 11:20
  clip for a YouTube intro. It is completely unpolished.
• 11:23
  The speaker is actively pointing left and right.
• 11:26
  But they are just pointing at empty space in
• 11:28
  the raw footage. Which looks totally bizarre
• 11:30
  without edits. Right. The edit has one main crucial
• 11:33
  job here. It needs to make those awkward gestures
• 11:36
  feel intentional and professional. We need to
• 11:39
  dynamically generate and place visuals exactly
• 11:42
  where the speaker points, perfectly timed to
• 11:44
  their hand movements. Step two in the practical
• 11:47
  workflow is writing the first goal prompt. You
• 11:50
  use the slash goal command in Codex. This keeps
• 11:53
  the agent heavily focused, working Strictly toward
• 11:55
  the finished result. You start by providing the
• 11:58
  exact file path of the raw video. And you explicitly
• 12:01
  tell it to transcribe, cut, and animate in that
• 12:05
  exact order. We talked about why that sequence
• 12:07
  matters earlier. You ask for a premium, highly
• 12:10
  polished finish. You request extremely fast pacing
• 12:14
  with almost zero dead space. Yeah, you ask for
• 12:16
  an Apple -style motion design aesthetic. Think
• 12:19
  smooth liquid glass, soft shadows, and incredibly
• 12:22
  clean animations. Then you describe the specific
• 12:25
  visual beats you want. On the left side, you
• 12:28
  ask for a translucent liquid glass card. You
• 12:30
  want animated charts to appear inside of it.
• 12:32
  On the right side, you ask for 3D animations
• 12:34
  with actual depth. You want spinning graphical
• 12:36
  elements and a subtle dark overlay to make it
• 12:39
  pop. For the subtitles, you request a standard
• 12:42
  lower third placement. You ask for crisp white
• 12:44
  text with bright blue word highlights as they
• 12:47
  are spoken. You even ask for previous video clips
• 12:49
  to play inside a 3D fan layout on the screen?
• 12:52
  It is a really detailed prompt. Very. You finish
• 12:55
  this massive prompt with a clear quality bar.
• 12:58
  You explicitly tell the agent to verify the entire
• 13:01
  edit with screenshots before presenting it. Step
• 13:04
  three is reviewing version one. The AI goes to
• 13:07
  work. The first result successfully reduced the
• 13:10
  video from 65 seconds down to a tight 28 seconds.
• 13:14
  The 3D animations and the core motion graphics
• 13:16
  were surprisingly strong. The transcript sync
• 13:19
  was mathematically perfect, but there were a
• 13:22
  few obvious glaring problems. Right. The opening
• 13:25
  two seconds felt a bit weak and empty. The 3D
• 13:28
  video cards occasionally clicked through each
• 13:30
  other in virtual space, and some of the informational
• 13:32
  text was just too small to read on a phone screen.
• 13:34
  Step four is giving targeted feedback to fix
• 13:37
  those issues. You write a second goal prompt
• 13:40
  focusing specifically on those problems. But
• 13:42
  here is an absolutely crucial prompting tip from
• 13:44
  the guide. You absolutely must tell the AI what
• 13:47
  worked well first. Right. If you do not tell
• 13:49
  it what worked, it might accidentally break the
• 13:51
  good parts. Yeah. It will try so hard to fix
• 13:54
  the bad parts that it regenerates the entire
• 13:56
  code base and ruins everything else you liked.
• 13:59
  Exactly. You anchor the successes first. The
• 14:01
  second version came back significantly better.
• 14:04
  It had a much deeper 3D layering on the cards.
• 14:07
  The text was scaled up and much more readable.
• 14:10
  Step 5 is testing a completely different opening
• 14:13
  sequence. Version 3 used a full -screen, highly
• 14:16
  animated cutaway. It had beautifully layered
• 14:19
  liquid glass cards taking up the whole frame.
• 14:22
  This really highlights the biggest advantage
• 14:24
  of AI video editing. Testing a completely different
• 14:27
  creative direction is incredibly fast. You are
• 14:30
  not spending four hours rebuilding an After Effects
• 14:32
  composition. You are just typing a new sentence.
• 14:35
  Step six involves making small, granular fixes
• 14:38
  manually. Yeah, you do not need to re -prompt
• 14:41
  the AI for every single tiny change. You can
• 14:43
  just use the localhost hyperframe studio interface.
• 14:46
  If the text is slightly too small, you don't
• 14:48
  ask the AI to fix it. You just open the studio
• 14:50
  and raise the font size to 80 manually. Because
• 14:53
  it is all HTML -coded, updates instantly. It
• 14:55
  is vastly faster than generating a full AI revision.
• 14:58
  By mixing plain England prompts with manual font
• 15:01
  tweaks, are we actually saving time or just shifting
• 15:04
  the labor? Oh, you are definitely saving massive
• 15:06
  amounts of time. The AI builds the entire complex
• 15:10
  3D architecture from scratch. You are just tweaking
• 15:13
  a tiny slider at the very end to perfect it.
• 15:16
  Right. The heavy lifting is automated. You only
• 15:18
  step in for the final polish. Exactly. Step 7
• 15:21
  is the ultimate payoff for this entire process.
• 15:25
  You turn the final, perfect edit into a reusable
• 15:27
  skill. Once the result looks exactly how you
• 15:30
  want it... You just ask the codex agent a simple
• 15:32
  command. Turn this into a skill. The AI analyzes
• 15:35
  all the code, the pacing, and the design rules
• 15:38
  and saves them. So the next time you edit a video,
• 15:40
  your instruction is tiny. You just say, edit
• 15:42
  this video using this skill. Every single round
• 15:45
  of feedback you gave becomes a permanent part
• 15:48
  of that skill. The AI reuses your preferred pacing,
• 15:51
  your exact font sizes, and your motion design
• 15:53
  on all future projects. Let's step back and look
• 15:56
  at the bigger picture. The shift happening here
• 15:58
  is really profound for the industry. It truly
• 16:00
  is. Video editing is fundamentally moving away
• 16:03
  from manually dragging individual clips on a
• 16:06
  timeline. Yeah, you are essentially acting as
• 16:08
  a creative director now, rather than a technician.
• 16:11
  You simply dictate the logic, the pace, and the
• 16:13
  visual style you want. Codex and Astra handle
• 16:16
  the intense brain work. HyperFrames builds the
• 16:19
  complex motion graphics. Eleven Labs handles
• 16:21
  the precise audio timing. The ultimate value
• 16:24
  here is not just creating one fast flashy video.
• 16:27
  It is the compounding value of saving your specific
• 16:31
  aesthetic. You build a reusable skill that gets
• 16:33
  better and more refined over time. Your personal
• 16:36
  brand's visual identity becomes a deployable,
• 16:39
  scalable asset. It is incredibly powerful for
• 16:42
  independent creators and small businesses who
• 16:44
  don't have massive post -production teams. Absolutely.
• 16:46
  I want to leave you with the final thought to
• 16:48
  chew on before we wrap up. two sec silence. If
• 16:51
  an AI can perfectly mimic your specific editing
• 16:54
  style and pace based on a simple text file of
• 16:57
  rules, how long until viewers can choose to watch
• 17:00
  any video on the internet automatically re -edited
• 17:02
  into their own preferred custom pacing right
• 17:05
  on the fly? Wow, that is a staggering thought.
• 17:08
  a completely customized viewing experience generated
• 17:12
  in real time for every single person based on
• 17:15
  their attention span. It completely changes the
• 17:17
  definition of what a final cut even means. Try
• 17:19
  running a short 60 -second clip through this
• 17:21
  workflow yourself. It is the best way to learn.
• 17:24
  Feel that shift from timeline editor to creative
• 17:27
  director firsthand. Thank you for joining us
• 17:29
  on this deep dive.