Your best coworker answers the same questions every week, and every answer evaporates the moment they log off. AI skill files fix that. They turn how your smartest colleague works into a file any AI assistant can load, follow, and reuse. Think of it as bottling judgment instead of just saving documents. The practice is spreading fast, and it is easier than it sounds.
What an AI skill file actually is
An AI skill file is a folder with one markdown file inside, usually called SKILL.md, that teaches an AI assistant how to do one specific job your way. That is the whole format. The file starts with a couple of lines of metadata, a name and a description of when to use it, and the rest is plain instructions written like you would explain the task to a new hire.
The format comes from Anthropic, which open-sourced the Agent Skills standard in late 2025. Here is the part that made it catch on: it is not locked to Claude. OpenAI adopted the same format in ChatGPT, where Skills now ship with Business and Enterprise plans, and VS Code supports it for coding agents too. A skill written for one assistant generally works in another without changes. Write once, run anywhere is finally true in the least glamorous, most useful way. Anthropic’s engineering post documents the full spec if you want to go deeper.
Skills differ from the custom assistants you may have tried. A Gemini Gem or a custom GPT is a whole persona you chat with. A skill is more like a procedure card. Your assistant keeps working normally, but when a task matches the skill’s description, it loads the file and follows the steps. We covered the developer-side version of this when Anthropic’s Skills API went GA. This guide is about the human side, and it works because AI skill files capture what a person knows, not just what a company documents.
Why bottling expertise caught on
Every team has someone whose work seems like magic. One editor’s newsletters never need revisions. One analyst’s spreadsheets never break. The freelancer whose client onboarding just works is the one everybody fights to hire. Their process lives in their head, and it walks out the door with them.
Dan Shipper’s newsletter Every has been tracking this trend all summer, with pieces titled “The Case for Cloning Your Coworkers” and “Turning your best colleagues into reusable skills.” The idea hit a nerve because it flips the usual AI anxiety. Instead of AI replacing people, it preserves what the best people do so everyone else can operate at that level. Companies in China are reportedly requiring employees to document their workflows for exactly this purpose, which kicked off a lively debate about who owns a person’s methods, the worker or the company. Every’s writeup on reusable skills is worth your time if this tension interests you.
There is a money angle too. A well-built skill is an asset. Freelancers package their methodology and deliver work faster. Consultants onboard clients quicker. Small businesses stop losing institutional knowledge every time someone quits. One repeatable process, captured once, saves hours every single week.
How to capture a coworker’s process in an afternoon
You need about three hours, your AI assistant of choice, and a colleague willing to talk. Coffee helps.
Step 1: pick one repeatable task
The best first target is a task the expert does the same way every time. Editing a client newsletter. Screening job applicants. Prepping a weekly sales report. If the process changes every time, save it for attempt number three, not number one. One AI skill file should do one job.
Step 2: interview the expert
Sit with them, or record a voice call, and walk through the task twice. Once for the happy path, once for the weird stuff. The questions that matter:
- What do you check first, and why that one?
- What mistakes do beginners always make here?
- When do you break your own rules?
- What does “done” look like? What does “done badly” look like?
- Show me the last three times you did this. What differed?
Record their answers. Those edge cases and decision rules are the actual gold. Anyone can write “proofread the text.” Only Sarah knows that the client’s legal team rejects any sentence starting with “ensuring,” so you flag those before sending.
Step 3: write the SKILL.md file
Now for the fun part. Open a new file, and put this at the very top, replacing the placeholder text:
---
name: newsletter-editing-sarah
description: Edits client newsletters following Sarah's checklist. Use when editing, reviewing, or polishing client newsletter drafts before sending.
---
# Newsletter editing process
1. Read the whole draft once without changing anything. Note the flow.
2. Check every claim has a source link. Flag any that don't.
3. Flag sentences starting with "ensuring" for legal review.
...
Then write the instructions in plain language, numbered steps, with the expert’s rules and two or three real examples. Good example plus bad example beats paragraphs of explanation. Keep the AI skill file under about 500 lines; if it balloons, split reference material into extra files in the folder and mention them by name.
The description line deserves real care. It is the trigger. A vague description means your AI never loads the skill when it should. Say what the skill is for and when to use it, in the same words a colleague would use when asking for the task.
Step 4: test it against the real thing
Here is the verification step most people skip, and it is the one that separates a useful skill from a horoscope. Take a task the expert already completed. Run the same raw input through your assistant with the skill loaded. Compare the output to the expert’s actual work. Where they differ, ask the expert what the file missed, and add the missing rule. Two or three rounds usually gets you to “this is genuinely her work.”
Step 5: share and version it
In ChatGPT, Skills can be shared with your workspace straight from the Skills page, so your whole team inherits the process. For Claude and coding agents, drop the folder into your project or share the file itself. Add a version number and a “last updated” line at the bottom. Processes drift. Without versions, you end up with three copies of the same skill, each slightly wrong, which is how rituals become superstitions.
One caution from Anthropic’s own docs: review any skill before installing one you did not write, especially skills that fetch things from the internet. A skill is instructions, and instructions can hide nasty things. Never put client names, passwords, or anything confidential into AI skill files you plan to share around.
Where skills work (and where they don’t)
Skills shine when the work has real structure. Editing, QA checklists, report generation, onboarding sequences, code review, brand-voice enforcement. Anything with rules and edge cases that a person currently applies from memory. AI skill files are at their best when there is a checklist hiding in someone’s head.
They disappoint when the work is genuinely improvisational, or when the “expertise” is really taste that nobody can articulate. If your interview in Step 2 produces shrugs instead of rules, you have learned something useful: that knowledge is not ready to be bottled. Come back after the next few times they do the task, and ask why each decision went the way it did.
Also worth knowing: a skill captures process, not presence. It will not make the intern a senior editor. It makes the intern useful on day one, which is worth a lot and oversold by nobody with a mortgage. We covered the developer side of this same idea when Skills shipped for Claude Code, and our piece on creating custom AI assistants pairs well with this one if you want the persona route too.
Try it with your own brain first
Before you recruit a colleague, bottle something you do on autopilot. You are an expert at something your AI currently handles clumsily, and you already know the hidden rules. AI skill files reward exactly that kind of honesty, and after you ship one, the interview method above gets much easier because you have felt where the gaps hide.
Capture one process this week. Your future self, and your replacement someday, will both be grateful.
