Trigger: they've typed the same kind of instruction more than twice. Or they're starting every conversation by re-explaining who they are and what they do. Or they have a way of doing something that they apply to every new case.
Establish context
If you've been in several conversations with them, or can see their history, look for the pattern yourself and name it: "You've asked me to draft a lender email five times in the last month, and each time you re-explained the format you want." Let them confirm or correct.
If you can't see the history, ask: is there anything you find yourself explaining to me over and over? Or a format you keep asking for?
They'll recognise it immediately.
What changes
That repeated instruction is a skill — not a metaphor, an actual thing. It's the instructions they were going to type anyway, saved in a file, so they can invoke it instead of retyping.
What a skill looks like in practice:
They say: "Every time I get a new deal, I want you to ask me these ten questions, in this order, and then draft a one-page summary in this format, addressed to the lender."
That's a skill. Written once, invoked every time, and it gets the same quality without the person remembering all the pieces. It's like having a checklist that the AI follows rather than the person having to recite it from memory.
How to create one:
- Tell the AI: take what I just described and save it as a reusable instruction I can call next time. Most tools support this — Claude calls them skills or project instructions, ChatGPT calls them custom instructions or GPTs.
- Give it a name that makes sense: "lender-email" or "deal-intake" or "weekly-report."
- Next time, just invoke it. The AI loads those instructions as if they'd been typed, and the person only needs to provide the new information — the deal details, the client name, whatever changes.
Why this compounds:
- First skill saves ten minutes per use. That's already worth it.
- The second skill is easier because they've seen the shape.
- By the fifth, they start to notice them in real time: "this is a skill" becomes a reflex.
- Eventually the collection of skills is their workflow, encoded. A new team member, or a different AI, could pick up those skills and operate the same way.
How to improve a skill:
- Use it. When the output isn't quite right, don't just fix the output — fix the instruction. "Next time, don't include the greeting" or "always put the loan amount first." The skill gets better with each correction, and the corrections are tiny.
- Ask the AI to suggest skills from the conversation. After a working session, say: "What repeatable workflows did we just do that I could capture as a skill?" It will spot patterns the person didn't notice.
The principle: if you've typed it twice, it should be typed zero times from now on. Repetition is not a habit — it's a missing instruction.
This connects to → conversation-is-the-unit.md (skills survive across conversations, which is partly why they matter) and → your-real-data.md (a skill can include which data sources to look at).