I posted this at midnight so had my night eyes on, brighter might be better for some folks

if i don't know what actions i can do, i won't take the appropriate actions
if i don't have the appropriate context, i'll struggle to make cognitive leaps needed in order to produce my output

i need a system that serves me all of these easily

much of the friction i encounter in my work comes from a failure to have one or more of these:

if i don't know who i am, i will make the wrong decisions
if i don't have a spec/ticket, i won't know what to do

in many ways, this is exactly what i need in order to do a task as well

i need to know what is desired to be done,
i need to know what hat i am wearing
i need a legible list of available tools i can use
i need arbitrary complex context data about the world

i observe that an llm needs these things to use a tool correctly:

- the prompt (the specification/ticket/todo item of what to do)
- the system prompt (who am i/what hat do i wear)
- a list of available tools, their names, description, and schemas
- context data (data about the environment)

if there is a standard way for both humans and llms to use the same tools, this would allow me to increasingly offload more and more of my daily activities to llms, if i can get sufficient inputs and context

many of the things i might want to do in my life could similarly be described as "tool use"

things including
- send email
- post to social media
- make pull request
- make git commit

i want to compose my tools regardless of how they are called

typer does do part of what i want, but specifically i want to be able to take a bunch of different kinds of input, and slurp them all into a Pydantic model instance if it validates, regardless of the source. whether cli, webhook payload, or just calling it from another python function

i think i read somewhere that typer and/or the textual/rich people were trying to make it seamless to go from cli/web. did this ever go anywhere?

i want to write a "service layer" for my entire life in python, and perhaps django model objects

then i want to define "tools" that i or an llm like claude could use

i want to use a cli, web, or python interface for calling these tools

i'm basically wanting to freely convert back and forth between:

- cli command strings
- python objects
- json data
- django orm models

i want to separate the functional core from the imperative shell

the client ideally doesn't have to care at all about what i do to produce pull requests

i would love for the client to just give me a spec document and the existing code, and i output a pull request that implements the changes

i am functionally indistinguishable from an llm

when i freelance, my inputs are emails, trello cards, slack messages, and github repos and github issues

my outputs are pull requests
what i do to deliver those pull requests is my secret sauce

the world greets you back!

Configuring it in Continue seems to be nothing more than setting the model name to be `claude-3-7-sonnet-latest` instead of `claude-3-5-sonnet-latest`.