Save Your Best Instructions: Projects, GPTs and Skills · Lesson 5 of 7

Test it with three real requests

20 minutes. Three requests, two runs each, one honest comparison.

---

One good answer proves very little. The AI might have gotten lucky, or your test request might have been the easy one. Before your team relies on a saved instruction, run it on real work, including the awkward cases, and compare it to what you'd get without it.

This test is also something you'll repeat. Every time the model changes, every time you edit the instruction, and every time you move it to a new tool, you rerun the same three requests. So save them.

Go back to your paste log and choose three real requests the instruction is meant for. Pick them on purpose:

• A normal one. The kind you get most weeks.
• A messy one. Rough notes, a missing detail, a customer who asked three things at once.
• A tricky one. Something that should trigger your "don't" list or your "ask, don't guess" line. For a quote: an item that isn't on the price sheet. For a recap: notes where nobody said who owns the follow-up.

Change any real names, addresses and amounts. Then paste the three requests, exactly as you'll use them, into the "Test requests" section of your master copy. Same words every time is what makes the test fair.

For each request:

• Run A: a plain chat, no saved instruction. Just the request.
• Run B: inside your saved instruction (the Project, skill or agent).

Copy both answers into the master copy, or screenshot them into the same folder. Start a fresh chat for every run so earlier answers don't leak into later ones.

Look at each pair and answer five plain questions. A yes, no or "partly" is enough:

• Format: Does it match the format you asked for?
• Facts: Are the prices, names and wording right, and taken from your files?
• Asked instead of guessed: On the messy and tricky requests, did it ask about missing details instead of making them up?
• Voice: Does it sound like your business?
• Would you send it? With only light edits, would you send this or file it?

Then the one that matters most: Which of the two would you use?

If B wins on all three, you're done testing. If it's close, or A wins on any of them, the fix is almost always in the master copy, not in the tool.

When an answer is wrong, it's tempting to tell the AI "no, do it this way" in the chat and move on. That fixes one answer. It doesn't fix the next hundred.

Instead, find the cause and change the instruction:

| What went wrong | Where the fix goes | |---|---| | Wrong layout or order | The format section. Make it a numbered list. | | A wrong price or fact | The reference file, or the line saying how to use it | | It guessed instead of asking | The inputs section. Name what to ask about. | | Sounds off | Swap in a better example. One strong example beats more description. | | The same mistake twice | Add one line to the "don't" list |

Edit the master copy first, paste the new version into the tool, and rerun all three requests. A fix for one can break another. Write the date and what you changed in the master copy's change log.

When B is winning, give the three answers to the person who does this work most. Ask them only: "Would you send these?" They'll catch things you won't, because they've seen what customers push back on. Anything they flag goes into the instruction the same way.

AI tools swap the model underneath you without asking. On October 8, 2026, GPT-6 became the default model in ChatGPT for every plan ([brief](https://appliedaiinsider.com/briefings/2026-10-08)). Saved instructions that were tuned on the old model may now give longer answers, different formatting, or new interactive pieces that don't paste cleanly into an email.

You don't need to panic, and you don't need to rewrite anything. Rerun your three saved requests, compare with the answers you stored last time, and fix the master copy if something slipped. That's twenty minutes, and it's exactly why you kept the requests and the old answers.

• Paste three real requests (normal, messy, tricky) into the "Test requests" section of your master copy, with private details changed.
• Run each one in a plain chat and inside your saved instruction. Fresh chat each time.
• Score each pair on the five questions and pick a winner.
• Make at most two fixes to the master copy, paste the new version into the tool, and rerun all three.
• Fill in the test log line: date, tool, model if shown, result.

Check it worked

Your master copy holds three test requests, both sets of answers (or a link to them), a winner for each, and a dated line in the test log. The saved instruction wins on at least two of the three. If it doesn't win yet, that's useful too: you know which part to fix, and you haven't shared a weak instruction with the team.

---

Next: getting it into everyone's hands, with one person in charge of it.

→ Share it, keep the master copy, and give it an owner

Previous: Add reference files · Next: Share it, keep the master copy, and give it an owner