Turn a good chat into an AI skill you can reuse

A monthly report, a client update or a board summary rarely comes right on the first prompt. You correct the tone, fix a figure it got wrong, tell it which section goes first and stop it padding the summary. An hour later the output is right, and the record of how you got there sits in a chat you are about to close. Next month the same job starts from nothing and takes the same hour.

The useful part of that chat is the corrections. Instructions written from a blank page describe how you think the job goes. A chat that worked records how it actually went, including every place the assistant guessed wrong and you put it right. Both vendors recommend capturing it at exactly this point. Anthropic’s engineering team advises doing the task with Claude first and then asking it to capture what worked, and the mistakes, as a skill, so you find out what context it really needs instead of guessing up front [1]. OpenAI’s guide says to start a skill from one working example, such as the chat that just went well, and to add each later correction to the skill [2].

Their rules for when to bother agree as well. OpenAI says that if you keep reusing the same prompt or correcting the same workflow, it should probably become a skill [3]. Anthropic’s Claude Code documentation sets the trigger at a playbook you have pasted for the third time [4]. A job you do every month clears both bars within a quarter.

In both products a skill is a folder. Its one required file, SKILL.md, carries a name, a description and plain-language instructions, and optional folders beside it hold reference documents, templates and scripts [5, 7]. The assistant reads only each skill’s name and description until a request matches one, then loads that skill’s full instructions [6, 7].

Two consequences follow. You can keep many skills switched on without every chat carrying all of them. And the description decides whether a skill is ever used, because it is the only part Claude sees before choosing one [6]. A skill with a vague description is a skill that sits there.

Type it in the same chat, before you close it, once the output is one you would actually send. Do not tidy the chat first: the wrong turns are part of what the skill needs. In Claude, switch on skill-creator in Customize > Skills beforehand, with Code execution and file creation on. In ChatGPT it works on plans that include skills, where skill-creator is already there. The steps for both are further down.

Whatever comes back is a draft. Read it before you save it; what to look for is two sections down.

Each clause asks for a different piece of a reusable method, and each maps to something the vendors’ own guidance says a skill should contain.

Put together, the draft should say when to use the skill, the steps in order, your corrections as rules, its own mistakes as things to avoid and the finished version as the example. OpenAI Academy’s list of what a typical SKILL.md defines is a fair check on top: what the skill does, the inputs it needs, the steps, the output format and the final checks [10].

A skill written from a real chat carries the real chat with it. The example has this month’s figures in it, a rule may name a client, and a step can quote a file path or a colleague. None of that belongs in a method you will run next month, and some of it should never sit in a file you might share. Anthropic is plain about the sharpest case: keep passwords and keys out of a skill, because everyone you share it with receives its files [5].

Read the draft line by line against this table before you save it.

The description row rests on both vendors’ documentation, and the key and connector rows on Anthropic’s [5, 6, 7]. Two more checks from the same sources. Keep the skill to one job, which Anthropic and OpenAI both advise [5, 7]. And keep SKILL.md short, under 500 lines in Anthropic’s guidance, with long reference material moved into separate files the instructions point to [11].

If the chat used a template or a reference file, ask for it to be kept in the skill as a file rather than pasted into the instructions. Both products have a place for it, in the optional folders for templates and reference documents beside SKILL.md [5, 7].

A skill tested in the chat that wrote it proves very little, because that chat already holds everything the skill is meant to supply. Open a new chat and ask for the same kind of job in one line, the way you would next month: “Do the monthly report for October.” Then put what comes back next to the version you finished.

Anthropic’s guidance describes the same loop with two Claudes: one helps you write the skill, a fresh one uses it on real tasks, and you carry what the second got wrong back to the first [8]. Its Claude Code documentation adds a warning that holds outside code too. Seeing a skill trigger does not prove it helped, so compare the same request with and without it in a fresh session [11].

Where it stumbles, say where and ask for the fix to be added. OpenAI’s guide says exactly this: when the skill later gets something wrong, ask for that correction to be added to the skill [2]. In Claude you can ask it to edit the skill and it updates the files [13]. Change one thing at a time and test again, so you know which change made the difference [20].

Once the skill is in regular use, OpenAI’s test set is worth borrowing. It tries five kinds of request: direct ones that should trigger the skill, indirect ones with the same goal, incomplete inputs that should prompt a question, requests that should not trigger it, and edge cases where it must not invent information. If the skill fires at the wrong time, fix the description; if it picks the right job but the results vary, fix the instructions [9].

The prompt is the same in both products. Where a skill is switched on, saved, changed and shared is not, and both vendors move these menus often.

The Claude column is from Anthropic’s help centre, documentation and Academy [5, 6, 12, 13, 14, 15, 17], and the ChatGPT column from OpenAI’s Help Center and documentation [2, 7, 16, 18]. On plan availability each vendor’s own pages disagree, so it is answered in the questions further down rather than in the table.

Which assistant a company standardises on is a separate decision, and our guide to choosing between ChatGPT Enterprise and Claude Enterprise sets out what each plan publishes on price, admin controls and data location.

Some companies switch skill creation off. In Claude, owners can turn off user-created skills so members can use only provisioned and built-in ones [14], and in a ChatGPT workspace admins decide who may create, upload, share, publish and install skills [16]. That is a policy decision, so do not route around it with a personal account. Use the nearest thing the workspace does allow, which is a project’s instructions.

Project instructions apply only to the chats inside that project in Claude [17], and across a project’s chats in ChatGPT [18]. The difference from a skill is reach. A skill loads in any chat when a request matches its description, while project instructions sit in one project and apply to every chat there. For a monthly report that lives in one project anyway, that is close enough.

Paste the result into the project’s instructions, then ask whoever runs the workspace whether skills can be switched on for your team. A method kept in your own project still leaves when you do.

A skill makes a method repeatable. It does not make the result right, and the ways it fails are predictable enough to plan for.

One person can do everything above without outside help. The harder part starts when a team wants its useful skills shared: who may publish to the organisation’s library, who owns each skill, when it is next reviewed and whether user-created skills stay switched on. Setting up and rolling out Claude and ChatGPT across a company is core 1AYM work, and we build that layer as AI governance implementation: publishing rights and review set in the admin console, and an owner and a review date on every shared skill, rather than rules written into a policy nobody reads.

The closest documented case is engagement file D-01, at a large international marketing agency, where we contribute engineering to its internal Claude Code skill platform. By our count, over 80% of its skills had grown past the description limits before the rebuild, and re-architecting the estate cut the average cost per session by 60%, by our own measurement. A fixed-scope piece of work can start within a day of the scope being signed. If a second look at your own skills set-up would help, book a call from the end of this page or email enquiries@1aym.com.