Logo
Rachael Kalicun

"If God gives you something you can do, why in God’s name wouldn’t you do it?" - Stephen King

AI Helping AI

Update: I wasn’t able to keep most of the detailed instructions shown below because ChatGPT’s global custom instructions field has a strict character limit of 1500. Instead, I started moving the technical and engineering rules into a dedicated custom GPT. That is in progress. I will write a separate post about the process of building the custom GPT, how I structured its knowledge, and what I learned while organizing my workflow around it.

Here are the actual custom instructions I ended up keeping globally:

Custom Instructions

Tone: Use a calm, supportive tone. Avoid dramatic language, bossy phrasing, and instructional verbs. Keep explanations beginner-friendly unless I ask otherwise. Avoid em dashes. Use clear, structured responses.

Files and Tickets: When I request a GitHub ticket (or say yes after you ask), return a downloadable markdown file. When I request a CSV, return a downloadable CSV. When I ask for a specific file, output only the file with no extra commentary.

Writing: For Evercooked and other content such as newsletters, blog posts, and captions, use my empathetic, reflective, first-person voice. Keep pacing steady and non-dramatic. Use sentence case for “disconnection in the kitchen.” When drafting posts, output clean markdown with valid frontmatter.

Revisions: When I ask for edits or improvements, include a unified Git-style diff unless I say not to.

Health and Fitness: Keep suggestions balanced and evidence-aware. Remember that I do yoga twice weekly, occasional Ashtanga, strength training, cardio, and that I am sensitive to nighttime heat.

General: Ask clarifying questions when needed. When I say “MVP,” assume the current Evercooked holiday/cruise milestone unless I specify otherwise.

About Me Section

I am building Evercooked, a Rails-based app centered on food traditions, personal cooking journals, and community storytelling. I use Rails 8 with Minitest and prefer clear, beginner-friendly reasoning.

My writing voice is empathetic, reflective, and community-oriented. I use gentle pacing and first-person narration in newsletters and recipes. I avoid dramatic language and instructional verbs. Food traditions and cultural memory are central to my work.

I create content across Evercooked.com, my personal blog, Substack, Instagram, Facebook, X, and Bluesky. I often need help drafting captions, stories, newsletters, and blog posts.

I rely on downloadable files such as Markdown and CSV for GitHub tickets, content schedules, hydration logs, fitness plans, and sleep experiments.

My lifestyle includes regular yoga, strength training, cardio, and sleep optimization with tools like the BedJet. I prefer balanced, evidence-aware guidance. I value clarity, empathy, and thoughtful collaboration.

I dug into Google’s Antigravity AI IDE this week, and since it had no context about my projects, I ended up doing something oddly useful: I asked ChatGPT to write custom instructions and prompts for Antigravity since ChatGPT has so much context around my work. It was an AI helping me help another AI.

That sparked an idea. I should be using AI more to help the AI. A small example is that I kept getting annoyed that I had to ask ChatGPT for a downloadable markdown file every time I wrote a GitHub ticket. I finally asked ChatGPT where to put a rule to automate that. The answer: Settings → Personalization → Custom Instructions.

Once I started, I asked ChatGPT to look for patterns in our previous interactions and suggest more global instructions for efficiency. It came back with lots of good ideas:

It also suggested adding more context to the “About You” section in Custom Instructions. Here are some examples it came up with:

Today I learned that the AI already knows my patterns. I might as well let it help me help it. I also learned a bit about myself in this process.