I spent the last few weeks doing something a little unusual: writing a cookbook. Not a real one. A prompt cookbook, twenty six tested, reusable prompts for ChatGPT, Claude, and Gemini, organized the same way you'd organize a recipe book, by what you're actually trying to make.
Here's the part that might actually be useful to you as a builder, not just a reader.
The problem I kept running into
Every time I wanted a model to do something specific and repeatable, a Socratic tutor for studying, a structured research synthesizer, a system prompt builder for my own agents, I'd rewrite the same prompt from scratch, slightly worse than the last time. No version control on my own thinking. No single source of truth.
So I built one. Not a prompt library buried in a Notion doc nobody opens. Something structured enough to hand directly to a model as its own instructions.
What that actually looks like
Each recipe follows the same shape: a role, a tool recommendation, and a prompt formula with the variables marked out in brackets.
You are a Socratic tutor teaching [subject or topic] to a [grade level] student.
Never give the direct answer. Instead, ask one guiding question at a time that
leads the student toward the answer themselves. Wait for their response before
asking the next question. If they get stuck twice in a row, offer a small hint
framed as another question, not a statement.
That's one of twenty six. Others cover research synthesis, system prompt design for persistent agents, debugging assumptions with explicit step by step reasoning, and a handful of less technical ones, meal planning, budget review, client onboarding, because the same discipline applies whether you're prompting for code review or a bedtime story.
The part I think is actually novel
Along with the book itself, I built three ready made skill files, one for ChatGPT custom instructions, one formatted as an actual Claude Skill with proper frontmatter, one for a Gemini Gem. Drop one in, and that model already has all twenty six recipes as reference material. Ask it in plain language for what you need, and it pulls the right formula instead of you re-deriving it every time.
If you've built internal tooling around reusable prompt libraries before, I'd genuinely like to hear how you structured it. This was my first pass at treating prompts like something closer to a dependency than a one-off message.
The full thing, book plus all three skill files, is here if you want to see the whole set: https://merakislove.gumroad.com/l/cookbook
Curious what other builders are doing for prompt reuse across models. Drop it in the comments.
Top comments (0)