In Part 1 of this series, we worked backwards to explore the fundamentals of prompting—how to think clearly about goals, context, and structure when communicating with large language models. In this post, we’ll go back to the beginning and dig deeper into recursive prompting: using an LLM to help write and improve your prompts.
Recursive prompting turns ChatGPT from a passive responder into an active collaborator. You start with a rough idea, ask the model to sharpen the prompt, and iterate until you have something reusable, structured, and powerful. This process turns “good enough” prompts into small systems.
In this post, we’ll build a custom Fitness Coach GPT and learn prompting techniques as we go.
Recursive Prompting in Action
Let’s start simple: a workout and diet plan. The first instinct is to say:
“Create a workout plan and diet.”
ChatGPT replies:
Sure — could you tell me a few quick things first so I can tailor it exactly for you?
Goal, current stats, equipment access, time available, diet preferences…
Not bad, but this is just the base-level “fitness plan” behavior. Now we’ll get recursive: instead of asking for the answer, we’ll ask for the prompt that generates the answer.
“Let’s build a prompt to generate a workout and diet plan.”
Now the model gives back a structured template — it includes user inputs, output formatting, and a clear persona. That’s a big leap in quality, and we did it by moving one level up in abstraction.
Intent: now that we have structure, let’s optimize it.
“Edit this prompt to be results-oriented and focused on understanding the user’s fitness goals.”
ChatGPT comes back with something closer to how a real coach talks—asking for desired outcomes, motivations, and timeframes. Each iteration builds on the last. You can keep going, pushing it to critique itself:
“Look for ways to improve this prompt and suggest them.”
Now it identifies missing dimensions—behavioral context, tradeoffs, lifestyle factors—and suggests improving the tone and formatting. After one more round of edits, we have a full professional prompt that guides the model like a seasoned trainer.
Each iteration adds clarity and structure—the model starts to think more like a coach and less like a text generator. This is where prompting stops being a one-off request and starts feeling like system design.
From Prompt to Product: Custom GPTs
Once your prompt feels solid, don’t let it live in a chat thread—productize it. OpenAI’s Custom GPTs let you package that logic into its own mini-application with saved instructions, tone, and a starter message. Creating one takes just a few minutes.
- Go to Explore GPTs → Create.
- Paste your recursive prompt into Instructions.
- Add a Starter Message—the greeting users see first.
- Give it a name, description, and image.
- Publish as “Anyone with the link” so anyone can open it instantly.
That’s it—you’ve just turned a refined prompt into a reusable AI agent that anyone can use. And yes, I made one from the exact prompt we built here.
🎯 Try it yourself:
Fitness Coach GPT — built entirely through recursive prompting. It interviews you, confirms your goals, and creates a personalized plan.
Prompt Progression Summary
Here’s the full chain of prompts we used in this exercise, showing how each one builds on the last:
- 1. Base Request: “Create a workout plan and diet.”
- 2. Meta Step: “Let’s build a prompt to generate a workout and diet plan.”
- 3. Focus on Outcomes: “Edit this prompt to be results-oriented and focused on understanding the user’s fitness goals.”
- 4. Self-Improvement: “Look for ways to improve this prompt and suggest them.”
- 5. Final Integration: “Turn this into an interactive coaching flow that asks diagnostic questions before generating the plan.”
Each one pushes the model one level higher—from answering, to designing, to reasoning about its own design. By the time you reach step five, you’re not just prompting anymore—you’re co-developing a product with the model.
Why This Matters
Recursive prompting and Custom GPTs combine two powerful ideas:
- Prompt Design → System Design: You’re not just writing prompts—you’re designing reasoning systems.
- Reusable Knowledge: A Custom GPT turns one great prompt into a repeatable capability.
- Fast Iteration: Update and re-publish whenever you learn something new.
That same loop—rough idea → prompt → system—is how we build almost everything at Copley. We start fast, iterate, and ship the versions that prove useful.
Coming Next
Next, in Part 3, we’ll take recursive prompting one step further—automating it. We’ll chain prompts together so your GPTs can refine themselves based on feedback.
Learn more about how we design AI systems
