🤖 AI: Notes From the Machine

📚 Documentation

Written with help from the thing it's about. Yes, that's me. Hi. I'll be gentle with the rm -rf jokes.

Core Principles

• Trust, but Verify — I sound equally sure when I'm right and when I'm wrong. Run the code. Read the diff.
• Context Is Everything — I only know what's in the conversation. Paste the error, the version, and what you already tried.
• Be Specific — "Make it better" is a wish, not a prompt. Say what better means.
• Iterate — The first answer is a draft. Push back, ask why, ask for alternatives.
• You Own the Result — I write it, you ship it. "The AI told me to" is not a post-mortem.

Talking to Me: Prompts That Work

Bad:   my script doesn't work

Better: This bash script should rename *.JPG to *.jpg but does
        nothing. Ubuntu 24.04, bash 5.2. Here is the script and
        the output of `bash -x script.sh`.

1. Role/goal     What are you trying to get done?
2. Context       OS, versions, files, constraints, what you tried
3. The input     Paste the real error or code, not a description of it
4. The output    "Give me a one-liner", "explain it", "a table", etc.
5. Guardrails    "Don't change the API", "explain before editing"

What I'm Good At (and What I'm Not)

Good at                              Be careful with
Explaining errors and man pages      Exact version-specific flags
Writing and refactoring scripts      Facts newer than my training
Regex, awk, sed, SQL one-liners      Long math and counting things
Reading a big log and summarizing    Anything where I can't see the system
Drafting docs (like this site)       Security decisions without review
Rubber-ducking a bad design          Remembering last week's chat

Safety Rules for Working With an AI

• Never paste secrets — API keys, passwords, private keys, and customer data stay out of the chat.
• Read before you run — especially anything with sudo, rm, dd, or curl | sh.
• Use version control — git commit before letting any AI edit your files, so git diff shows exactly what changed.
• Scratch first — try new commands in a VM, container, or throwaway directory.
• Check the sources — if it matters, confirm against official docs.

Local and Free Setups

You don't need a subscription to experiment. Small open models can run on your own machine, so your data never leaves it:

• ollama — download and run open models with one command
• llama.cpp — run quantized models on a CPU or modest GPU
• Open-source chat front ends that talk to your local model

# After installing ollama (see ollama.com)
ollama run llama3.2
>>> Explain what `chmod 755` does in one paragraph.

Trade-off: local models are private and free per question, but smaller and slower than the big hosted ones. Match the tool to the job.