Skip to main content
Skills are reusable instructions that tell an AI client how to approach a recurring workflow, which Control MCP tools to prefer, and how to verify the result. They do not grant additional access: the connected user, workspace, session access level, and enabled Control features still determine which tools are available.

​
Before you add a skill

  1. Connect the client to Control MCP.
  2. Refresh the client’s tool list.
  3. Confirm that the tools referenced by the skill are available to the current user.
  4. Keep tenant-specific IDs, thresholds, and reporting choices outside shared skill files.
The MCP tools/list response is the source of truth. Update a sample when the tool names or schemas shown by your client differ from the example.
If your client supports skill files, copy an example into a file named SKILL.md in the location documented by that client. Otherwise, save the same text as project instructions or paste it at the start of a workflow.

​
Analysis and reporting

Financial variance review

Compare periods and investigate the largest supported movements.

Monthly management report

Create a reconciled monthly report or editable board-deck brief.

Budget versus actual

Align a plan with actuals and explain material variances.

Entity performance comparison

Compare explicitly named legal entities on a consistent basis.

Cash-flow review

Review consolidated cash movements and investigate material changes.

Transaction investigation

Turn an account, vendor, or unusual movement into bounded evidence.

Vendor overlap review

Review counterparty spend, flag overlapping services, and quantify savings.

Uniform invoice review

Analyze normalized invoices and drill into invoice-line evidence.

Stripe journal-entry review

Reconcile Stripe revenue, fees, cash, disputes, VAT, and FX.

​
Context, planning, and controlled writes

Financial context pack

Give an AI client reviewed context about layouts, accounts, dimensions, and currencies.

Excel financial model

Build an editable workbook from reconciled Control actuals and plan data.

Budget import

Validate a structured workbook and preview budget writes before committing them.

Manual adjustment proposal

Investigate, propose, confirm, and verify manual ledger adjustments.

​
Run a review on a schedule in Control

When Agents is enabled for your workspace, several of these reviews are also available as hosted workflow templates. A hosted workflow runs on Control after the month is ready or on a schedule, without an AI client staying open, and it can only use the tools its administrator granted. Hosted templates stay generic. Put your own account rules, recurring monthly steps, and materiality thresholds in Agents > Workspace guide instead of editing a template. The ledger review stops and asks for the guide when none exists, and the other templates use it when it is present. Organization admins maintain the guide, and connected AI clients can read it with the get_agent_workspace_guide tool. Hosted runs read data in bounded pages and have a per-run tool budget. When a review covers more data than one run allows, the run is marked as failed and incomplete. When the model can summarize what it gathered, the failure message includes that partial result and states what was and was not covered; otherwise it only says the run was incomplete.

​
Adapt a sample safely

  • Add customer-specific reporting choices only after they have been reviewed.
  • Keep credentials, tokens, personal data, and destination secrets out of the skill.
  • Start with read-only workflows and require human review before consequential writes.
  • Test the skill against a known period and reconcile the result with Control before using it routinely.
  • Treat generated presentations and spreadsheets as editable outputs, not as a second source of truth.