vim / ai

I type <Leader>r from a Markdown file in Vim to send the contents to an LLM such as o1 and display the output in a vertical split:

I can embed files from my codebase into my prompt because the Markdown is passed through mdembed before it is sent to the LLM:

I can type <Leader>c to continue the last conversation.

Vim config

In laptop.sh:

# https://github.com/croaky/laptop/blob/main/laptop.sh

# AI via CLI
go install github.com/charmbracelet/mods@latest
go install github.com/croaky/mdembed@latest

In init.lua:

-- https://github.com/croaky/laptop/blob/main/vim/init.lua

-- Helper functions
local function buf_map(bufnr, mode, lhs, rhs, opts)
  opts = vim.tbl_extend("force", { noremap = true, silent = true, buffer = bufnr }, opts or {})
  vim.keymap.set(mode, lhs, rhs, opts)
end

local function filetype_autocmd(ft, callback)
  vim.api.nvim_create_autocmd("FileType", {
    pattern = ft,
    callback = callback,
  })
end

local function run_file(key, cmd_template, split_cmd)
  local cmd = cmd_template:gsub("%%", vim.fn.expand("%:p"))
  buf_map(0, "n", key, function()
    vim.cmd(split_cmd)
    vim.cmd("terminal " .. cmd)
  end)
end

-- Markdown
filetype_autocmd("markdown", function()
  run_file("<Leader>r", "cat % | mdembed | mods", "vsplit")
  run_file("<Leader>c", "cat % | mdembed | mods -C", "vsplit")
end)

LLM config

Configure mods with your preferred LLM:

mods --settings

Example:

default-model: o1
apis:
  openai:
    base-url: https://api.openai.com/v1
    api-key:
    api-key-env: OPENAI_API_KEY
    models: # https://platform.openai.com/docs/models
      o1-preview:
        aliases: ["o1"]
        max-input-chars: 792000
        fallback: o1-mini
      o1-mini:
        max-input-chars: 500000
  anthropic:
    base-url: https://api.anthropic.com/v1
    api-key:
    api-key-env: ANTHROPIC_API_KEY
    models: # https://docs.anthropic.com/en/docs/about-claude/models
      claude-3-5-sonnet-20240620:
        aliases: ["claude", "sonnet"]
        max-input-chars: 680000
  ollama:
    base-url: http://localhost:11434/api
    models: # https://ollama.com/library
      "deepseek-r1:latest":
        aliases: ["deepseek"]
        max-input-chars: 500000

To set up Ollama models such as DeepSeek-R1:

brew install ollama
ollama run deepseek-r1

So what?

I like this analogy:

The lower the gear in a car, the more control you have over the engine but you can go with less speed. If you feel in control, go to a higher gear. If you are overwhelmed or stuck, go to a lower gear. AI assisted coding is all about grokking when you need to gain more granular control and when you need to let go of control to move faster. Higher level gears leave more room for errors and trust issues.

When I want to work in a lower gear, I use small chunks of code in my Markdown prompt. When I want to work in a higher gear, I embed one or more files in my Markdown prompt.

The best results I've had using LLMs are when I prepare a good, isolated prompt. The longer a conversation goes back-and-forth with follow-ups, the more the LLM gets confused and won't produce what I want.

I prefer to start with a fresh prompt for each request (<Leader>r), editing my tmp.md based on anything I learned from earlier LLM requests, embedding files from the codebase (some of which I may have edited).