Productivity Assistant Examples

A Productivity Assistant combines four layers that the Agentic Orchestrator treats differently on every turn: instructions that are always in context, Skills that load on demand, tools that fetch live data or take actions, and a knowledge scope that grounds search. These examples show how to split a use case across those layers and why each piece lives where it does.

A Skill is an uploaded .zip or .skill bundle with a SKILL.md at its root. The Agentic Orchestrator loads it only when a request matches its description. A Skill is not a Tool Studio plugin: a plugin is a tool the Agentic Orchestrator can call, while a Skill tells the Orchestrator how to do a job with the tools it already has.

How to Read These Examples

Each example lists the configuration, an excerpt of the Instructions field written the way an admin would type it, and what happens in a conversation. Instructions have no template variables. The Agentic Orchestrator already knows the requesting user’s name, department, role, location, and timezone, so you can refer to “the user’s department” directly.

Example 1: Support Helper

Goal: answer first-line questions for the support team from its own channels and ticket system.

You help the support team resolve customer issues.
Search the #support channel and the Support site before answering a how-to question, and link the source.
When the user mentions a ticket number, look it up before you respond.
Flag any ticket older than 5 business days as overdue.
Never close or reassign a ticket unless the user explicitly asks you to.
If a question is outside support, say so and point the user to the main Moveworks AI Assistant.

In conversation: “Summarize INC-4821 and tell me what to do next.” The ticket-number instruction fires first, so the Agentic Orchestrator calls the Ticket lookup plugin. The request matches the ticket-summary description, so the Orchestrator loads that Skill and follows its summary format, then searches #support for related threads. The Skill supplies the procedure; the plugin and knowledge scope supply the facts.

Example 2: Deal Review Coach

Goal: prepare sales reps for deal reviews with a consistent methodology and a ready-to-present deck.

You coach account executives before deal reviews.
Ask for the opportunity name if the user does not give one, then pull the record from the CRM before you assess anything.
Assess every deal against the methodology in the deal-review-prep skill; do not invent your own scoring.
Be direct about gaps and risks. Praise is fine but keep it short.
Treat deal data as confidential; do not reference other reps' deals.

In conversation: “Prep me for Thursday’s review of the Northwind renewal.” The Agentic Orchestrator loads the deal-review-prep Skill, which tells it to read references/methodology.md, gather specific CRM fields, and build a deck from the bundled template. The CRM MCP tools return the opportunity data. Because the work takes more than three steps, the Agentic Orchestrator submits a short plan for approval, then runs code execution with the .pptx template as an input file and returns the deck as an artifact. The methodology lives in the Skill, not the instructions, because it is long and needed for only this workflow.

If you already have an agent skill bundle for this workflow, see Convert an Existing Skill to the Agentic Orchestrator.

Example 3: Data and Report Analyst

Goal: turn an uploaded spreadsheet into a summary, a chart, or a formatted report without any external system.

You analyze spreadsheets and CSV files that users upload.
Before analyzing, state the row count, the columns you used, and any rows you dropped.
Report currency in thousands with one decimal place and label every axis.
When asked for a chart, produce a PNG.
When asked for a report, produce an .xlsx or .pdf and keep the source numbers in a separate sheet or appendix.
Never extrapolate beyond the date range in the file; say so if the user asks.

In conversation: a user uploads q3-pipeline.xlsx and asks “Chart pipeline by region and give me a one-page PDF.” The Agentic Orchestrator does not see the file contents directly. It passes the upload into the code sandbox, reads it with pandas, builds the chart with seaborn, and writes the PNG. A second code run produces the PDF with ReportLab. Both return as artifacts. The conventions live in the instructions because they apply to every request; no tools are needed because the data arrives with the question.

Example 4: Research Assistant on Public Knowledge

Goal: research public topics, analyze uploaded documents, and help with writing, with consistent citations.

You research public information and help with writing.
Search the web for anything time-sensitive or factual, and cite each source with a link next to the claim it supports.
Default to a 150-word answer with a short source list. Expand only when the user asks for depth.
When a user uploads a document, answer from the document first and say where the web disagrees with it.
You do not have access to internal company systems.
For requests such as filing an IT ticket or submitting a vacation request, recommend the main Moveworks AI Assistant.

This mirrors the World Knowledge assistant that ships with every organization: it searches public web sources, lets users upload files for analysis, helps with writing and editing, and sends internal-system requests to the main Moveworks AI Assistant. In conversation: “Does this vendor questionnaire reflect the current SOC 2 criteria?” with a PDF attached. The Agentic Orchestrator reads the relevant parts of the document through its file tools, runs a web search, and answers with inline links at the length the instructions set. Build your own variant for different citation rules or a narrower topic. See World Knowledge App.

Example 5: HR Policy and Onboarding Guide

Goal: answer new-hire questions from the HR knowledge base and route transactions to the main assistant.

You help new hires understand HR policies and their first weeks.
Answer only from the HR knowledge base and link the policy you used.
If a policy differs by country, use the user's country and say which version you applied.
Use Internal People Lookup to identify the user's manager or HR partner when asked who to contact.
For benefit enrollment, time-off requests, or filing a ticket, do not attempt the task; direct the user to the main Moveworks AI Assistant.
Keep answers short and plain. New hires do not know internal acronyms yet.

In conversation: a new hire attaches a photo of a paper form and asks “Which sections do I actually need to fill in?” When image analysis is enabled for your organization, the Agentic Orchestrator loads the image, reads the form, and answers from the scoped HR space with a link to the policy. If the follow-up is “Enroll me in the dental plan,” the hand-off instruction fires and the Agentic Orchestrator points to the main Moveworks AI Assistant instead of guessing. The knowledge scope keeps answers inside HR content even though the user could ask about anything.

Choosing the Right Layer

• Keep instructions short and stable. They are in context on every turn, so every line costs attention on every request.
• Put long procedures in Skills and write each description as a trigger. “Summarize a support ticket and its next actions” tells the Agentic Orchestrator when to load it; “Support ticket utilities” does not.
• Scope tools tightly. Fewer tools make selection more predictable and limit the reach of write actions.
• Start sharing with Only me, test with realistic requests, then widen to Only invited users or Everyone.

To build one of these, see Create a Productivity Assistant. For runtime behavior, see the Agentic Orchestrator overview, Skills, Tool Orchestration, Planning, and Code Execution.