The hockey stickPhase 2 · Tinkering

System Prompts That Run on Repeat

You have all had this happen: after fifteen minutes of back-and-forth, AI finally gives you a great output. A system prompt is how you get that output in one step next time, instead of fifteen minutes every time.

4 min read

What a system prompt is

Think of it as a recipe. It has the ingredients and the steps, and the order matters. Same ingredients in the wrong sequence give you mush; in the right sequence, a cake. A system prompt is the saved set of instructions that produces a specific kind of output reliably, so you do not rebuild it from scratch.

The difference from a master prompt: a master prompt is context about you. A system prompt is a repeatable job ("write our weekly customer update," "turn a call transcript into a proposal").

How to build one

  1. 1
    Use pull prompting to write it. Say: "You are an expert AI engineer. Build me a system prompt that does [the job]. Ask me every question you need first."
  2. 2
    Answer the questions with voice-to-text.
  3. 3
    Test it. Run the prompt and look at the output.
  4. 4
    Refine in plain English. "Tweak this part," "make it shorter," "always include X." You are coding the prompt with words.
  5. 5
    Save it where it runs on repeat. Paste the finished system prompt into a custom GPT (ChatGPT), a Project (Claude), or a Gem (Gemini). Now anyone can give it the topic and get a consistent result with no instructions.

Why this is the bridge to Wiring

When a prompt lives inside a saved GPT, Project, or Gem that your team can run by dropping in a topic, you have left improvising behind. That is exactly the Wiring phase: saved prompts, a project or two, the first things that run without you re-explaining. This resource is the on-ramp.

Curious how far this goes? There was a public leak of the system prompts behind tools like Perplexity, Notion, and Lovable. Searching "system prompts of AI tools GitHub" turns up the collection, a useful look at how serious these can get.

From Auto-Phil

Auto-Phil helps owners capture the great output that usually takes fifteen minutes of back-and-forth and get it in one step. The company sets up system prompts that run on repeat, so the good result is the default rather than the reward for persistence.

Keep going

When you want a hand

Skip the guesswork on your own setup.

Thirty minutes, no pitch. Tell us the work you do and we will tell you the next move that actually fits your shop.

Free Coaching CallBack to The Tinkering Library