### AI Workflow Setup

- Demonstrated Granola meeting transcription tool
  - Records and generates notes automatically
  - Doesn’t store transcripts, only outputs notes
  - Used for every work meeting including sensitive discussions
- Introduced Cursor IDE as advanced coding platform
  - AI-powered development environment (recently acquired for ~$60B)
  - Replaced VS Code with integrated AI assistance
  - Command K for inline AI editing and formatting
  - Terminal integration for system commands

### Image Generation Automation System

- What we currently do (manual):
  - Each game week, 8 short videos are made using photos of key NRL players as reference.
  - Videos are kept to 3–5 seconds each to save money and make it easier to redo if something looks wrong. The best clips are combined into a final edit.
  - Editing is done in DaVinci Resolve, including manually cutting players out from backgrounds to place them over text and graphics.
- Making Claude Code skills (mostly automatic, with a human checking key steps):
  1. The AI searches the internet for recent photos of each player automatically — no manual downloading needed.
  2. The AI checks that the player in the generated image actually looks like the right person before moving on.
  3. Images are generated in batches. The AI checks each one against the brief and automatically tries again if something looks off.
  4. Once images are approved, the AI turns them into short video clips, writing tailored descriptions for each scene.
  5. Sound is added using a tool called elevenlabs — if the video's built-in audio is good enough, it's kept; otherwise, sound effects or a voiceover are layered on top.
  6. All files are automatically named and organised in a logical way so nothing gets lost.
  7. The AI flags any images that might use copyrighted photos and leaves those for a human to decide on.

### How It's All Organised

- Everything lives in a simple folder on your computer — no complicated setup needed
  - There's a main instruction file (think of it like a "house rules" document) that tells the AI what it's doing, which tools to use, and how to handle common problems.
  - Each part of the process — images, video, audio, editing — has its own subfolder with its own specific instructions.
  - Saved shortcuts ("skills") can be reused and chained together, e.g. "make images → check them → name them → turn into video."
- Tools being used:
  - **Fal.ai** — the main image generation service; you pay per image.
  - **OpenAI (GPT-2 image model, released two days ago - search the web)** — another image/video option depending on the task.
  - **Eleven Labs** — generates voiceovers and sound effects.
  - **Automated web browsing** — the AI can search the internet on your behalf to find player photos without you lifting a finger.

### Making It Fast and Affordable

- What's been working:
  - Satirical NRL content has hit 3M+ views — the rough, inconsistent look actually works in its favour and makes it feel more authentic.
- The main pain points:
  - Generating a full batch of images can take 40+ minutes, so you don't want to get halfway through and realise the brief was wrong.
  - Longer videos cost roughly 4 times more to generate than short clips — a 40-second video vs. a 10-second one is a big difference in credit spend.
- How to keep costs down:
  1. Always work in short 3–5 second clips and combine them afterward rather than trying to generate one long video.
  2. The AI checks that the player's face looks right before committing to a full batch — this stops wasted runs.
  3. If a batch keeps failing, it stops automatically rather than running for 40 minutes and producing nothing usable.
  4. All files are named automatically in a consistent way so they're easy to find and sort through later.

### Smart Automation Features with Claude Code and skills

- Saved shortcuts ("skills") that do the heavy lifting:
  - A "30 day search" skill that searches the internet for the latest AI image and video prompting advice from the past 30 days — because this stuff changes fast.
  - A "image prompt improver" skill that takes a rough idea and rewrites it into a detailed, cinematic description that gets much better results from the AI.
  - Define context. The AI only reads the extra reference files it actually needs for a given task — so it doesn't slow down or get confused by irrelevant information.
- What Claude Code can do on your computer:
  - Create and organise folders, move files, fetch team logos, and keep a set of reference photos for each player — all without you doing it manually.
  - Browse the internet in headless mode the background to source images, even without a visible browser window open.
- Quality checks built in:
  - Every image is checked against the original brief before being sent.
  - Player faces and jersey numbers are verified for consistency - could reference a local .md with up to date numbers and names. 
  - Any time something goes wrong, every skill should log what happened to a .md with enough context (not too much, just enough, keep it concise) and then we can iterate on the skill in a new session and improve it. this ensures we aren't repeating the same errors and we can improver all the skills over time. 

### Partly completed steps

- Claude Code set up in the desktop app
  - Install the necessary system packages
  - Add your account keys for Fal.ai, OpenAI, and elevenlabs so the AI can access those services — these are stored privately on your machine.
  - Get a basic image generation test working end-to-end.
- Build the core skills
  - Set up the player photo sourcing and face-matching step.
  - Get batch image generation running with automatic naming and quality checks.
  - Define how many generated images are needed for each video (sometimes 1, sometimes 3, sometimes start and end frame)
  - Build the video generation step so approved generated images are turned into clips with a Kling 3.0 or Kling Omni video skill that uses fal.ai

### Next steps
- Expand into a full production pipeline with multiple chained skills
  - Add audio: voiceover and sound effects generated and mixed automatically.
  - Add a rough-cut editing step where a claude skill assembles the clips, with an option for you to approve or tweak the edit before export.
  - Add safeguards: rate-limit handling, and regular updates to keep the skills current as AI models evolve.