[Docs index](/docs.md) / [Agent Skills](/docs/agent-skills/overview.md) / Playbook: Build a Team AI Usage Cleanup Skill

---

# Playbook: Build a Team AI Usage Cleanup Skill

Your teams are already using AI. Support reps are summarizing escalations in their personal Claude accounts. A sales manager has a discovery-call prompt in a Google Doc that three people copied last quarter. Operations runs a weekly status update through ChatGPT every Friday. Finance has an invoice-review checklist nobody can find. None of it is wrong, but none of it is yours — the prompts live in personal accounts, the company has no way to review them, and nobody can answer "what AI workflows are we actually relying on?"

This playbook walks you through building an Agent Skill that helps an operations or enablement lead inventory that scattered AI usage, classify each workflow by risk and value, and produce a prioritized list of what should move into Assist first. The skill itself becomes the way you move from "AI is happening everywhere" to "we know what AI work the company depends on and we own the important parts."

## What you will build

By the end of this playbook, you will have:

- A workspace Agent Skill named **Team AI Usage Cleanup** with a `SKILL.md` procedure, a department intake checklist, a workflow inventory template, a customer-data policy, and a worked example.
- An approved version of that skill, mounted to an Operations or AI Enablement agent.
- A repeatable inventory process: an operator runs the agent against one department, gets a classified list of AI workflows, and decides what becomes an Agent Skill, what gets cleaned up first, and what stays out of Assist for now.
- A pattern you can run again for the next department without starting over.

## What you need before you start

- An Assist workspace where you can create workspace-scoped Agent Skills.
- A workspace admin who can review and approve the skill before rollout.
- An operational agent in your workspace that handles internal operations or AI enablement work (an `Operations Agent`, `AI Enablement Agent`, or similar). If you do not have one, [create a subagent](../subagents/creating-a-subagent.md) first.
- A first target department in mind — usually support, sales, operations, or customer success, because they tend to have repeatable written workflows and clear managers.
- One department lead who is willing to walk through their team's AI usage with you the first time you run the inventory.

## Step 1: Create the draft skill

Open Assist and go to **Workspace > Agent Skills**. Click **New Skill**. Set:

- **Name:** Team AI Usage Cleanup
- **Slug:** team-ai-usage-cleanup
- **Scope:** Workspace
- **Description:** Helps an operator inventory scattered AI usage across a department, classify each workflow by risk and value, and recommend which workflows should become managed Assist skills.

The skill is created as a draft. Nothing is attached to any agent yet, and nothing is live. Drafts are the safe place to work.

If your workspace prefers to start in chat, you can ask the AI to scaffold it instead:

> "Create a new workspace Agent Skill called Team AI Usage Cleanup. Slug: team-ai-usage-cleanup. It will help an operator review scattered AI usage in one department and produce a prioritized cleanup list. Start with a draft `SKILL.md` and supporting folders for `checklists`, `templates`, `policies`, and `examples`."

The AI creates the draft and the folder structure, and you can edit each file from there.

## Step 2: Author the main procedure

Open `/SKILL.md` and write the procedure the agent should follow. The first lines tell Assist when to use the skill — this is what the AI reads when it decides whether to pick this skill up:

> Use this skill when a user asks to inventory, review, or clean up scattered AI workflows before moving approved workflows into Assist.

Then walk through the actual process. A good `SKILL.md` for this skill includes:

1. Ask the operator which department they are inventorying and who the department lead is. Refuse to proceed without an owner — an inventory without a named owner just produces a document that gets ignored.
2. Collect existing AI workflows from the department: prompts in personal chats, prompt docs, Slack threads, spreadsheets, wikis, and informal automations. Use the intake checklist in `/checklists/intake-questions.md`.
3. For each workflow, gather: name, current location, who uses it, how often, what input data it touches, whether the output is internal or customer-facing, and whether it affects money, commitments, hiring, legal, security, or customer trust.
4. Classify each workflow using the categories below.
5. Apply the customer-data policy in `/policies/customer-data-rules.md` to flag workflows that touch sensitive data.
6. Produce the inventory in the format defined by `/templates/workflow-inventory.md`.
7. Recommend the first three workflows the department should standardize and the first one to retire.

The point of the procedure is to give the agent a path, not just permission. A vague skill produces a vague inventory.

## Step 3: Add the supporting files

The supporting files are what make the skill actually useful. The `SKILL.md` is the spine; these are the muscles. Create:

- `/checklists/intake-questions.md` — the questions to ask the department lead and operators. Keep it under 15 questions. Examples: "What prompt do you use most often?" "Where is it written down?" "What goes wrong when someone new joins the team and tries to use it?"
- `/templates/workflow-inventory.md` — a strict output format the agent must follow for each workflow: name, location, users, business value, risk, recommendation, next step. Strict formats produce reviewable inventories.
- `/policies/customer-data-rules.md` — what counts as sensitive: customer names, account identifiers, contract details, employee PII, financial figures. The agent uses this to flag workflows for redaction before standardization.
- `/examples/workflow-candidates.md` — three or four worked examples of workflows that should move into Assist, written in the inventory format. Examples teach the agent the bar better than instructions do.
- `/examples/do-not-migrate.md` — two or three examples of workflows that should stay manual or get retired. Equally important: it stops the agent from over-recommending.

You can author these files directly in Assist, or drive the work from chat:

> "Draft `/checklists/intake-questions.md` for the Team AI Usage Cleanup skill. Twelve questions an operator should ask a department lead during an AI usage inventory. Cover what prompts they use, where the prompts live, who else uses them, whether they touch customer data, and whether the output is customer-facing."

Review every file before submitting. The AI is good at first drafts; you are the one who knows which questions actually matter for your company.

## Step 4: Use the classifications that drive action

A mid-sized company does not need a theoretical AI audit. It needs a list with verbs on it. Add this classification table to the `/templates/workflow-inventory.md` template so every workflow lands in one bucket:

| Class | Meaning |
|-------|---------|
| Move to Assist | Repeated, useful, suitable for a reviewed Agent Skill |
| Needs cleanup first | Useful but contains private data, unclear instructions, or missing review rules |
| Keep manual | Too sensitive, rare, or judgment-heavy for an agent right now |
| Duplicate | Overlaps with another candidate and should be consolidated |
| Retire | Outdated, low-value, or no longer aligned with team practice |

Five buckets, five clear next steps. The operator can leave the inventory meeting with assignments, not impressions.

## Step 5: Submit for review and approve

Once the draft files are written, submit the version for review. From the skill detail page, click **Submit for review**. The version moves into a pending state.

A workspace admin opens the pending version and walks through it: the `SKILL.md`, the checklist, the template, the policy, the examples. The reviewer should be asking three things:

- Is the procedure specific enough that two different operators would produce comparable inventories?
- Does the customer-data policy match how the company actually defines sensitive data?
- Are the examples honest about what should and should not move into Assist?

When the admin approves, the version becomes available to mount. If the admin rejects, the feedback comes back as comments on the version, and you author a new draft addressing the points.

## Step 6: Attach the skill to the right agent

Open your target operational agent (Operations Agent, AI Enablement Agent, or whatever you named it) and go to its skills view. Click **Attach skill**, pick **Team AI Usage Cleanup**, choose the approved version, and set the mount path:

`/skills/team-ai-usage-cleanup`

Avoid vague mount paths like `/skills/v2` or `/skills/cleanup-final`. The mount path becomes part of the agent's working context — it should be readable a year from now.

The attachment pins the agent to this specific approved version. The agent will keep using it until someone explicitly upgrades the mount, which is exactly what you want. A reviewed workflow should not silently swap underneath the team.

## Step 7: Run the first inventory with a real department

Now use the agent. Pick the first department and sit with the department lead the first time the agent runs. A few real prompts to start with:

> "Inventory the support team's AI usage. The lead is Priya. We want a list of every recurring prompt or workflow the team uses, classified by what should move into Assist first."

> "Walk me through the prompts the sales team is using for discovery-call summaries. There are at least four versions floating around. I want a classified inventory and a recommendation for which one to standardize."

> "Review the operations team's weekly reporting workflow. It currently runs in a personal ChatGPT account. Classify it and recommend whether it should become an Assist Agent Skill."

The agent walks through the procedure: it asks the intake questions, collects evidence, classifies each workflow, applies the customer-data policy, and produces the inventory in the format the template defines. The output should be specific enough that the department lead can leave the meeting with a list of next actions.

If the agent over-recommends moving everything into Assist, push back in chat: "Be stricter. Only recommend 'Move to Assist' for workflows with more than two recurring users and a clear owner." That feedback should go into the next draft of the skill, not just this conversation.

## Step 8: Improve the skill from real inventories

After two or three real inventories, the gaps show up. Maybe the agent missed a class of workflow (informal Slack-bot scripts), or the classifications collapse two different problems together, or the customer-data policy is too narrow. Open a new draft of the skill, fix the files, submit for review, and have the admin approve a new version.

Then deliberately upgrade the agent's attachment to the new version. Existing mounts stay pinned to the old version — that is a feature. The team sees that the skill changes are reviewed and announced, not silent.

## What you built

You have a workspace Agent Skill that turns "AI is happening everywhere on this team" into a reviewable, repeatable inventory. The skill has:

- A specific procedure for one operator to run against one department.
- A checklist of questions that produce comparable inventories across teams.
- A policy that flags sensitive workflows before anyone tries to standardize them.
- A strict output format that gives a department lead a list with verbs.
- A worked example of what should and should not move into Assist.

More importantly, you have a way to make this kind of work shared infrastructure. The next operator does not invent their own AI-usage audit; they use the approved skill. The next department lead does not see a new format; they see the one the company decided on. The work is owned by your organization instead of by whoever wrote the best prompt doc this quarter.

This is what bringing scattered AI use under one platform looks like in practice: you do not start by banning anything. You start by knowing what exists.

## Where to go from here

- **Standardize the top candidates.** Take the "Move to Assist" workflows from the first inventory and build them as their own Agent Skills. The [Team Prompt Standardization](playbook-prompt-standardization.md) playbook walks through this.
- **Run the rollout deliberately.** Once you have approved workflows to attach to real operational agents, use the [Department AI Rollout](playbook-department-ai-rollout.md) playbook to keep the attachments controlled.
- **Add a quarterly cadence.** Schedule the cleanup agent to run a fresh inventory each quarter for the departments where AI adoption is moving fastest. The point is not to police usage; it is to keep the inventory current.
- **Track the metrics that matter.** Count workflows moved to Assist, duplicate prompts consolidated, risky prompts caught before rollout, and team members who stopped maintaining their own private prompt copies.

## Related guides

- [Creating a managed AI workflow skill](creating-a-managed-ai-workflow-skill.md)
- [Authoring and approving Agent Skills](authoring-and-approving-agent-skills.md)
- [Attaching skills to operational agents](attaching-skills-to-operational-agents.md)
- [Governing Agent Skill versions](governing-agent-skill-versions.md)
- [Troubleshooting](troubleshooting.md)

---

## Navigation

### In this section: Agent Skills

- [Agent Skills](/docs/agent-skills/overview.md)
- [Use Cases and Playbooks](/docs/agent-skills/use-cases.md)
- [Troubleshooting Agent Skills](/docs/agent-skills/troubleshooting.md)
- [Attaching Agent Skills to Operational Agents](/docs/agent-skills/attaching-skills-to-operational-agents.md)
- [Authoring and Approving Agent Skills](/docs/agent-skills/authoring-and-approving-agent-skills.md)
- [Creating a Managed AI Workflow Skill](/docs/agent-skills/creating-a-managed-ai-workflow-skill.md)
- [Governing Agent Skill Versions](/docs/agent-skills/governing-agent-skill-versions.md)
- [Importing a Team AI Workflow Skill](/docs/agent-skills/importing-a-team-ai-workflow-skill.md)

#### Playbooks

- [Playbook: Build a Department AI Rollout Skill](/docs/agent-skills/playbook-department-ai-rollout.md)
- **Playbook: Build a Team AI Usage Cleanup Skill** (current)
- [Playbook: Build a Team Prompt Standardization Skill](/docs/agent-skills/playbook-prompt-standardization.md)

### Other sections

- [MCP Servers](/docs/mcp-servers/overview.md)
- [Tool Creation](/docs/tool-creation/overview.md)
- [Agent Filesystem](/docs/agent-filesystem/overview.md)
- [Chat Sharing](/docs/chat-sharing/overview.md)
- [Scheduled Triggers](/docs/scheduled-triggers/overview.md)
- [Sandcastles](/docs/sandcastles/overview.md)
- [Subagents](/docs/subagents/overview.md)
- [Workspace Permissions](/docs/workspace-permissions/overview.md)

[Back to docs index](/docs.md)