Your team already uses Copilot, Claude, ChatGPT, or Gemini. A skill factory is one repeatable way to capture the prompts that work and turn them into skills the whole company can run. A small committee decides which ones deserve real development time, and AutomationTactics sets the system up with you in 90 days.
People in your company are already using AI on their own. A project manager writes a prompt that turns a messy field report into a clean daily log. A biller works out how to get a clean claim summary. These prompts work, and they save real time.
The problem is where that work lives. It sits in one person's chat history. Nobody else can find it. Two desks over, someone solves the same task from scratch and gets a different result.
When that person leaves, the prompt leaves too, and the next hire starts over.
A skill is a written, reusable set of instructions for one repeatable task, saved so an AI assistant does it the same way every time. A skill factory is the company-level system that produces those skills on a schedule, keeps the good ones, and puts development time where it pays off.
It runs in four stages.
Every department names the repeatable tasks worth turning into skills.
Employees build skills in the AI tool they already use.
Skills that need more than a chat tool get put forward for help.
A small committee decides what gets built, held, declined, or funded.
Ideas do not show up on their own. You have to ask for them on a set schedule. Once a month, send every department a short request. Name the tasks you do over and over that an AI assistant could do the same way each time.
Keep the ask small. Each submission covers one task, fills in a few fields, and takes about two minutes. A monthly rhythm keeps ideas coming without turning into a second job for anyone.
Each idea goes through a short intake form so you capture enough to sort it later. Score every idea on five things.
The scoring works the same way as the Pain Score in the AutomationTactics Automation Opportunity Analysis, which multiplies volume, frequency, error rate, and time cost, each rated 1 to 5.
The backlog lives in one place everyone can see, such as a list in SharePoint, Microsoft Lists, or a shared tracker. Give that list one owner.
Most people cannot list their repeatable tasks from memory. The AI tool they use every day already has a record. Copilot can see your email, calendar, and files, and Claude and ChatGPT keep memory and past chats.
Paste this prompt into the tool you use most and let it do the first pass. Then submit the best results through the intake form.
On Microsoft 365 with a Copilot license, this pass works especially well, because Work IQ lets Copilot read your mail, calendar, Teams messages, meetings, and files. There is a separate walkthrough for Copilot users with a prompt tuned to those sources.
I want to find tasks I do repeatedly that could become reusable AI skills. A skill is a written set of instructions for one repeatable task, saved so you do it the same way every time.
Look through everything you have access to about me and my work: memory, past conversations, files, email, calendar, and any connected systems. If you cannot access one of those sources, say so and skip it. Do not guess or make up tasks you have no evidence for.
From what you find, list up to ten tasks that meet all of these conditions:
1. I have done it, or asked you to help with it, more than once.
2. It follows roughly the same steps each time.
3. The output has a consistent shape, such as a report, a summary, an email, a checklist, or a spreadsheet.
For each task, give me:
- Task name
- The evidence you found (which chats, files, or messages point to it)
- How often it seems to happen
- The inputs I supply each time
- The output it produces
- A one-sentence description of when the skill should be used
- Data sensitivity: does the task touch client, financial, health, or personnel data
Rank the list by how often the task happens and how much time it seems to take. Put the strongest candidates first.
At the end, tell me which sources you were unable to check.Read the results as a starting point. The tool can only see what it has been shown, so it will miss tasks that happen in systems it cannot reach. Pick the two or three tasks you recognize as real and frequent, and submit those.
The employee who does the task builds the skill in the tool they already use. Every major tool has a place for it.
Microsoft 365 Copilot has Agent Builder, which Microsoft describes as an easy way to build declarative agents from a natural-language Describe tab or a manual Configure tab, plus Cowork Skills. Claude has Skills and a skill-creator. Anthropic launched Agent Skills on October 16, 2025 as folders of instructions, scripts, and resources that Claude loads when a task calls for them. It published the format as an open standard on December 18, 2025.
ChatGPT has Projects and custom GPTs. Gemini has Gems. The click-by-click steps for each are in the separate guide, How to Create a Skill.
Every skill contains the same seven parts, whatever tool it is built in. The parts are what make a skill reusable.
The sharing rule. A skill stays with the person who built it while it is still being tested, or while it only helps that one person's own work. It gets shared when it works on real tasks, a second person could use it, and it has passed a review against the seven parts above. Sharing means the skill is published to the company library with a named owner.
Most skills stay in the chat tool and need nothing more. Some need something the chat tool cannot do. Those are the ones to nominate for development help.
A skill is a candidate for development when one of these is true.
Who can nominate, and when. Any employee can nominate their own skill, or a department champion can nominate one on behalf of a teammate. A skill can be nominated once it has run for real work at least a handful of times and the person can show what it produces. A working skill with a real output makes a stronger nomination than an idea.
Nominations need a decision, and decisions need a small group with the authority to make them. This is the steering committee.
Keep it small so it can meet and decide.
The fuller AI steering committee described in the AutomationTactics FAQ adds legal or compliance and HR. For deciding on skills, a group of four keeps decisions moving.
How often it meets. Once a month is a good starting cadence. Right Partners, in its AI steering committee guidance, says the exact cadence can evolve and that monthly is a useful starting point for active adoption. Monthly is frequent enough that decisions do not pile up, and spaced enough that the group decides rather than reports.
For a nomination that cannot wait, circulate a written approval between meetings with a set response window. Governance guidance from firms such as AccuroAI and brics-econ describes the same practice, including a 14-day fast track for low-risk systems and electronic sign-off inside a 72-hour window for urgent initiatives.
What it decides. For each nomination the committee picks one of four outcomes.
What happens after a decision. Every outcome gets written down with a reason and a named owner. A build or a fund decision gets a start date and a check-in. A hold gets a revisit date. A decline gets a short note back to the person who nominated it so they know why and know their version still stands.
Score each nomination 1 to 3 on each criterion. Add the scores. Higher totals go first.
| Criterion | 1 | 2 | 3 |
|---|---|---|---|
| Frequency | Monthly or less | Weekly | Daily |
| Time saved per run | Under 15 minutes | 15 to 60 minutes | Over an hour |
| People who use it | One team | A few teams | Most of the company |
| Error cost when done wrong | Low | Noticeable | High or regulated |
| Effort to build | High | Medium | Low |
| Readiness | Task not yet documented | Partly documented | Documented and stable |
| Data sensitivity | Highly sensitive, needs controls | Some sensitive data | No sensitive data |
| Role | Responsibility |
|---|---|
| Employee | Submits ideas, builds skills for their own tasks, nominates the ones that outgrow the chat tool. |
| Department champion | Runs monthly idea generation for their area, helps teammates build skills, reviews skills before they are shared. |
| Skill owner | Owns one published skill, keeps it correct, answers questions about it, retires it when it is no longer used. |
| Committee member | Meets monthly, scores nominations, decides build, hold, decline, or fund. |
| Developer or AutomationTactics | Builds the skills the committee funds, handles system access and scheduling, hands the finished skill back with an owner named. |
| Weeks | What happens |
|---|---|
| 1 to 2 | Document how the work runs today and pick the pilot departments. |
| 3 to 6 | Train employees in the pilot departments to build skills in the tools you already own. |
| 7 to 10 | Run the first monthly idea and nomination cycle, and stand up the steering committee. |
| 11 to 13 | Build the first two or three nominated skills, then hand the process off to your team. |
A skill gets reviewed against the seven required parts before it is shared with the whole company. Every shared skill has one named owner. Once a quarter, each owner confirms their skill is still used and still produces the right result. A skill that is no longer used, or that no longer matches how the work is done, gets retired and removed from the library so nobody relies on something out of date.
Not at the start. Most skills are built by the people who do the work, inside the chat tool they already use. You only bring in a developer for the smaller set of skills the committee funds, the ones that need system access, scheduling, or multi-step automation.
Standardize on the tool you already pay for and already use. For most companies on Microsoft 365 that is Copilot. The skill factory works the same whichever tool you pick, because a skill is a written set of instructions before it is anything else.
That is fine at the idea and skill stages. The intake form, the seven required parts, the nomination process, and the committee are the same across tools. When you publish a shared skill, note which tool it was built in so the next person opens the right one.
Paste the idea prompt from Stage one into the AI tool you use most. It reads your past chats, files, and email where it has access, and lists the tasks you repeat. Submit the ones you recognize. Copilot users have a dedicated walkthrough.
A skill holds instructions. The data comes in when the person runs it. Write skills so the person supplies the specific data at that moment, and keep client names, health information, and financial records out of the saved instructions.
Sort data by sensitivity on the intake form, and route anything sensitive to the IT or security lead on the committee before the skill is shared. Microsoft Purview ships a default set of sensitivity labels to make this call consistent (Personal, Public, General, Confidential, and Highly Confidential), and defines Highly Confidential as your most sensitive data, shared only with named recipients.
The employee keeps using their chat-tool version. A decline only means the company will not spend development time on it right now. The skill still works, and it can be nominated again later if the task grows.
One meeting a month. If nominations come in well prepared and pre-scored, most can be decided in a few minutes each. Between meetings, a written approval handles anything urgent.
A prompt is something you type once. A skill is that prompt written down with a name, steps, inputs, an output format, and an example, saved so anyone can run it and get the same result.
In 90 days your company can have a working skill factory. Your people build the skills, and a committee decides which ones get development time.