¯\_(ツ)_/¯

Custom GPTs That Test Themselves

How I use self-simulated user testing to make custom GPTs smoother, faster, and more fun to use.

Illustration of a custom GPT running a self-simulated user test in a feedback loop

I've been building small custom GPTs I use every day — a few even took off after I added them to my portfolio. The biggest unlock was realizing: you can use a GPT to test itself — quickly, repeatedly, and without extra tools.

This turns iteration into a tight loop: simulate a user → watch friction → tweak → re‑run.

What This Is (and Why It Works)

  • What: Ask your GPT to act like a first‑time user, step through the flow, and critique the experience as it goes.
  • Why: You surface confusing branches, missing instructions, and tone issues before real users ever feel them.
  • Result: Faster polish, fewer regressions, and a smoother first impression.

How To: Run a Self‑Simulated User Test

  1. Paste this into your GPT:
You are now simulating a user test of this GPT.
- Walk through the flow as if you're a real user.
- Call out anything confusing or awkward as it happens.
- Try alternate branches and edge cases.
- After each run, pause and wait for my feedback before restarting.

  1. Watch the run. Note friction points (labels, choices, copy, missing guardrails).

  2. Tweak one thing. Keep changes small and intentional (wording, emoji affordances, ordering, safety rails).

  3. Re‑run. Confirm the fix. Repeat.

Example: Home Maintenance Assistant

I built a Home Maintenance Assistant for quick household fixes.

Mock Test Run (shortened)

GPT (as user): 👋 Hi there! What would you like to fix today? 1️⃣ Start a new maintenance issue 2️⃣ Continue a previous issue

Me (as builder): Feels a bit dry. Add clearer affordances with emojis.

GPT (self‑feedback): Noted. Adding visual cues could improve clarity. Restarting…

GPT (new version): 👋 Welcome! 1️⃣ Start a new maintenance issue 🛠️ 2️⃣ Continue a previous issue 📂 3️⃣ Change language 🌍

Me: Better. Now option 1 reads cleaner — continue and verify the next step.

Tips That Compound

  • Name the goal upfront: “In 2 steps, get the user to X.”
  • Design the menu: Use clear verbs, minimal choices, consistent emoji affordances.
  • Instrument copy: Prefer short, skimmable prompts over paragraphs.
  • Add safety rails: Ask clarifying questions before acting; confirm destructive steps.
  • Iterate in shots: Change one thing, re‑run, commit if the experience improved.

Why This Matters

Most first runs feel rough without feedback. This loop lets you explore branches early, spot edge cases, and upgrade the feel before anyone else touches it.

Try It Now

Paste the test block above into any of your GPTs and run one loop. Tweak one thing. Re‑run. You’ll feel the polish accelerate.

The best part: you already have everything inside ChatGPT. No extra tooling.

Check out more of my custom GPTs here.