Last Tuesday you asked for a quote follow-up letter from your site-visit notes, pushed back three times, and the fourth draft was the one you sent. This week you have another site visit and the chat from last Tuesday is forty sessions down the list. Copying the first prompt back in gets you the first draft again, the one you rejected.
A skill is the fix for that. It stores the method behind the good result, so you can attach it to a new chat and start from draft four.
Open the conversation, skip the first prompt and read what you typed after it, because the method is in those replies.
In the quote example, the follow-ups might have been “keep it under 250 words”, “put the price and what’s excluded in the first paragraph”, “don’t promise a start date”, and “sign off with my mobile, not the office line”. None of that was in the first prompt, yet it’s the reason draft four was the one you sent.
Copy each correction into a scratch note as a plain instruction. Add anything you fixed by hand afterwards, like a paragraph you always delete. Those edits are rules the AI never heard.
Some good results were one-offs. A skill pays its way when the same kind of input arrives again and the same kind of output should come back.
| Keep it a normal chat | Make it a skill |
|---|---|
| The input is different every time (a one-off rebrand, a single tricky email) | The input has a recognisable shape (site notes, a weekly export, a request email) |
| The output format changes with the audience | The output has a fixed shape (a letter, a table, a one-page summary) |
| You will do it once or twice | You will do it monthly or more |
| You can’t yet say what a good result looks like | You can list the rules a good result follows |
If it straddles both columns, run it as a chat a few more times until the rules settle.
A prompt asks for one thing. A skill describes how the thing is done by someone who wasn’t in last Tuesday’s chat.
The US Environmental Protection Agency’s guidance on standard operating procedures was written for labs and field crews, and it carries over well. It says an SOP should be detailed enough that someone with limited experience “can successfully reproduce the procedure when unsupervised”, and that it should be concise, step-by-step, in the active voice and present tense (EPA QA/G-6, April 2007). Read “someone with limited experience” as the AI on a fresh chat with none of your history.
A workable skill body has six parts. Here is the quote follow-up written that way:
# Quote follow-up letter Goal: turn site-visit notes into a follow-up letter that goes out with the quote. Inputs: the site-visit notes (typed or a photo of handwriting) and the quote PDF. Steps: 1. Read the quote total, the inclusions and the exclusions from the quote PDF. 2. Pull the client's name, the address and the two or three things they cared about from the notes. 3. Write the letter. First paragraph: the price and what is excluded. Second: how the work answers what they cared about. Third: how to accept. Rules: - Under 250 words. - Never state a start date. Say "we'll confirm a start date when you accept". - Use the price from the quote PDF only, never a figure from the notes. - Sign off with the mobile number in the notes, never the office line. Output: the letter as a document, ready to edit. Checks before finishing: - The total in the letter matches the quote PDF exactly. - Every exclusion in the quote appears in the letter. - If the quote PDF is missing, stop and ask for it.
The rules are your old corrections, now written as instructions. The checks are things the AI can verify against the inputs before it hands over, which is different from a worker’s finish line (more on that in how an AI agent knows it’s finished).
Inputs are where a second run tends to drift. The first time, you had the notes and the quote open. The second time, someone forgets the quote and the AI fills the gap with a plausible number.
So name every input in the skill and say what happens when one is missing. “If the quote PDF is missing, stop and ask for it” is one line, and it turns a confident wrong letter into a question. Do the same for a price list or contact sheet the skill relies on.
Examples are useful and also risky. La Trobe University Library’s prompting guide describes giving the tool several examples so it follows the logic. In a reusable skill, the AI may copy an example’s content as well as its logic, so a sample letter for a deck job leaks deck details into a bathroom quote.
Save examples for judgements the rules can’t capture, like one short pair showing how a vague client worry (“I don’t want it to look cheap”) becomes a sentence in the letter. Leave out the full sample letter. If you want reference material on hand, put it in a separate file inside the skill (for example references/past-letters.md) and name it in the steps as background reading for tone.
On desktop, open Settings from the bottom-left avatar, choose Skills, then New in editor. The Skill editor opens with a SKILL.md file in the Files list on the left.
references/past-letters.md. Duplicate and delete buttons sit above the content, along with undo and redo.The header also has Export (a .md file, or a .zip for a multi-file skill) and Import for a .skill, .md or .zip someone sent you.
The skill is still a draft until it has worked on something other than the job it came from. The EPA guidance makes the same point about procedures: it is “especially helpful” if a draft is tested by someone other than the person who wrote it. Here the stand-in for that person is a fresh chat with none of last Tuesday’s history.
Run it three ways:
Identical prompts can still produce different responses, as UCL Library Services’ prompting guide notes, so judge the pattern across all three runs. Every time it misses, fix the skill, not the letter. Add the rule, tighten the step, then run the same case again.
In any chat, click + in the composer, open the Skills tab and choose the skill. A labelled chip marked “skill” appears above the composer before you send. That chip is your confirmation the method is in the conversation. Without it, the reply is a normal chat reply. The same menu attaches brand kits and connectors (tutorial).
When the way you quote changes, open the skill from Settings, Skills (Open in editor), change the rule and rerun your three test cases. The EPA guidance suggests reviewing procedures periodically and withdrawing ones nobody follows. A skill list gets cluttered the same way, so deactivate the ones you no longer use. The Create your first reusable AI skill tutorial runs through these steps in five short lessons.
Build a skill from a chat that worked, test it on a second case, and attach it from the composer.
A saved prompt repeats the first message. A skill holds the method, inputs, rules and checks, and it is attached alongside whatever you type, so you can still add “this client is a repeat customer” in the message.
Long enough to hold every rule you’d otherwise retype. Past a page, move reference material into a separate file and keep SKILL.md to the method.
Use the chat as your source, but write the method yourself from your follow-up corrections and from any edits you made by hand after copying the draft out.
A skill tells a chat how to do a job. A worker goal tells an autonomous worker when the job is finished. They work together, and how an AI agent knows it’s finished covers the goal side.
Yes. Export it from the Skill editor header and send them the file. They open the Skill editor on their own account and use Import.
Lesson 7 of the Foundations course covers it briefly.