Personal AI Workflow

All this claude code / openclaw bullshit is very overwhelming to keep track of. So much noise on the internet about it. Still, I feel like people don’t discuss the actual specifics of what their workflow is. Getting this right for yourself is incredibly valuable right now. Having an effective workflow is the difference between being able to make progress on 10 things at once vs stalling due to inefficient context switching.

Using this as a scratchpad to muse on what an effective system would look like for me. Then I’ll look into current solutions.

First some criterion. A great personal AI has the following properties:

• Intelligence
• Persistent memory and recall
• Easy to context switch between multiple projects
• can do multiple things at once asynchronously
• shares context between everything

What is the hierarchy here?

• General
  • Projects
    • individual changes Is general life management considered a project? Probably not. What context would you want to keep for all new chats? I like the idea of ephemeral chats that don’t use any context and don’t pollute context for the future (like /btw in claude)

Let’s work through some use cases:

1. Create and manage a new project
   1. Website. Ex: “Who is this lyric from”
      1. local or remote, output would be testing link or video proof
   2. App. Ex: MagicMidi or Avalon app
      1. local, output would be app that i can actually use
   3. Extensions: Jailbreak website to be used by AI
      1. local, output would be extension to test
   4. Automation: settle up splitwise every week
      1. This is the only one that doesn’t have to have a local output
2. ML experiments
   1. has to be connected to cluster, but I personally don’t want to be connected to the cluster. should be able to launch an experiment from my phone.
   2. I want strong job monitoring, as well as reports on how the experiments turn out, and visibility into what the actual changes are so I don’t launch useless training jobs
   3. Also have to be able to test experiments easily, and eventually turn them into features
   4. It’s important for experiments to be rebased on updated code often.
3. Add a new feature to a system
   1. Should be able to asynchronously call the feature to be created. i.e. every new idea should come with an implementation.
      1. More important part is how can we review that implementation.
      2. low risk
         1. just need proof of it working
            1. Video proof
            2. Interactive demo that you can use
      3. high risk
         1. one of the above
         2. view code change
4. School work, i.e do this test
   1. should be able to operate the browser
   2. browser should be visible
   3. can use jailbreak hack
   4. has to be authenticated!
      1. Auth seems like a big blocker for making agents do real work on behalf of you. To do schoolwork, it has to be signed into LEARN. I think agent having access to your emails is quite strong
         1. the only other option is for the agent to have its own accounts / emails
         2. How can i make sure that my agent is logged into whatever i’m logged into?
         3. Should I let my agent make payments on my behalf?