How to Turn a Good Chat into a Reusable Skill

Anyone who just got a rare clean answer can turn a good chat into a reusable skill in 15 minutes: extract the trigger, the input, the output shape, and the stop, then throw away the instance. Do it in ChatGPT, Claude, or Gemini while the thread is still open, not next month when you only remember the vibe.

A good chat is an accident until you name the law.

If you copy the whole transcript into a note, you saved a fossil. Read what a skill prompt actually is if the difference is still fuzzy, then run the extract below.

Extract the loop. Delete the instance.

The instance is this week’s names, numbers, and paragraphs. The loop is what you would do to the next pile.

1. Mark the moment it worked

Scroll to the turn where the output became usable. Not the first prompt. Not the compliment at the end.

Write one sentence: “It worked when I forced X and forbade Y.”

Send that sentence to the same model: “Do not rewrite my sentence. List the instructions from this chat that produced X and forbade Y. Ignore the content about this week’s project.”

If Claude, ChatGPT, or Gemini starts summarizing the project, it is still hugging the instance. Cut the project names out of your sentence and ask again.

2. Fill the four-field card

Demand this shape, nothing else:

• Trigger — when you open the skill, in one line.
• Input — what you must paste, in one line.
• Output — the shape, the cap, the refuse token.
• Stop — what you will not ask next.

“Fill these four from the working turn. Under 120 words. No examples from this chat. If a field is missing, write UNKNOWN.”

UNKNOWN on the stop is common. People remember the prompt and forget the moment they should have shut up. You write the stop if the model cannot.

This is the same compression job as turning a messy note into a one-page brief. The chat is the dump. The card is the brief.

3. Strip names, dates, and lucky facts

Do a human pass with a red pen.

• Product names become “the three candidates I name.”
• Dates become “the week I paste.”
• The clever phrase the model coined becomes optional, or it dies.

Tell the model: “Replace every proper noun from this chat with a slot. If you cannot slot it, drop the line.”

A skill that still says “the Lisbon trip” is a souvenir.

4. Rerun once on a new, small input

New chat. Paste only the card. Paste a different input. Same job, different instance.

If the output shape holds, you have a skill. If the model asks for the old facts, the card still leaks the fossil. Fix the leak. Do not “add more examples.”

One rerun. Not a bake-off. You are testing the card, not hunting a better paragraph.

Store the card where you will open it on a bad Monday. The Promptcrates tools for reusable skills are built for that save. If you want to see how other loops are filed, browse the Promptcrates library and steal the card shape, not the prose.

5. Retire the original thread

Archive or forget the lucky chat. Leaving it pinned trains you to reopen a museum.

The next time the trigger fires, you run the card. You do not reread the Lisbon thread to remember how you felt.

If your card is longer than a page, it is still a chat.

Reusable skill card

Trigger: A chat just worked and you can feel yourself about to pin it “for later.”

Input: The working turn, your one-sentence “it worked when,” and a ban on this week’s proper nouns.

Output: A four-field card (trigger, input, output, stop) under 120 words, slots instead of names, UNKNOWN if the stop was never said.

Stop: After one rerun on a new input. No transcript dump. No third example. No pinning the fossil.

Takeaways

• The working turn is the source, not the whole thread.
• Four fields: trigger, input, output, stop. That is the skill.
• Strip names. A souvenir is not reusable.
• Rerun once on a new input. Luck fails that test.
• Archive the old chat. Open the card next time.