Today we are building an agentic workflow and let’s understand why this matters before we get into it.
If we want to extract real value from agentic AI, we have to think in workflows, not tasks. A task is a single step. A workflow is a connected sequence that runs end to end without you holding it together in between.
The example I am sharing today solves a pain point that shows up in every type of project for almost every type of knowledge work: the constant overhead of staying on top of what happened, what was decided, and who needs to hear from you. It can be time consuming when you are the one doing it manually.
I built this in Claude Cowork as an automated scheduled task - it fires every day at midday without me touching it. If you are on Microsoft 365 Copilot Frontier (Microsoft's early-access license for the latest AI features) with tasks enabled, you can build something similar there.
Here is the exact setup.
One caveat before we go further: this is not a static document. You will need to tweak it continuously to fit your workflow and your needs. I update mine often. I am still learning what works, what to add, and what to simplify. Think of it as a living system, not a finished product.
Let's get into it.
Ethan Mollick has argued that most companies are stuck automating legacy tasks instead of using AI to imagine radically new workflows. It's the equivalent of using electric light simply to extend working hours, rather than rethinking how work gets done entirely. - Ethan Mollick, Wharton professor, author of Co-Intelligence
Here is what that looks like in practice.
Below is the prompt template I use. I have this set up as a scheduled task in Claude Cowork to fire every day at midday without me touching it. In Claude Cowork, this prompt lives as a skill file - a plain text file that tells the scheduler what to run, when to run it, and what tools to connect. The prompt below is the core of that file. You can set it up per your preferences.
Start here to make it yours: paste it into your AI tool first and let it ask you questions. It should ask what tools you use, where your context files live, and how your setup is organized before it does anything. If it just runs without asking, add this line at the top: "Before you begin, ask me clarifying questions about my tools and context files so you can run this correctly."
What you need before you start:
A call recorder: Granola, Fireflies, Otter, or Zoom AI. Something that produces notes from your meetings.
A context file: one document describing who you are, your active projects, and your current priorities. A Notion page, Google Doc, or plain text file all work.
A contacts or stakeholder file: a running doc of the people you meet with: name, company, role, anything relevant. This is what makes the AI route tasks correctly and write follow-up emails that feel personal.
A PM tool with projects already set up: ClickUp, Notion, Asana, Todoist. Organize it so the project names map to your clients or work areas.
You are my project manager. Run my post-meeting sync.
STEP 0: Load context before doing anything else.
Read [path to your context file: a doc, Notion page, or text file
where you describe your work, current priorities, and key relationships].Read [path to your contacts/stakeholder file: CRM export, Notion contacts
page, or a running notes doc with what you know about the people you meet].Use this context to interpret everything below. Do not start fresh.
HANDLING RULE: If you cannot match a person to your contacts file or a task
to a known project, do NOT guess and do NOT create the task.
Flag all unmatched items in the output report under "Needs your attention."
I will decide whether a new project needs to be created or where it belongs.
STEP 1: Read today's meeting notes from [your call recorder: Granola / Fireflies / Otter / Zoom AI summary].
STEP 2: For each meeting, extract:
Key decisions made
Action items (with owner, rough due date, and which project it belongs to)
Anything to watch or follow up on
Any new information about the people I spoke with
STEP 3: Create tasks in [your PM tool: Notion / Asana / ClickUp / Todoist] for each
action item assigned to me — but only if you can match it to a known project.
If you cannot match an action item to a known project, do NOT create the task.
Flag it in the output report under "Needs your attention" instead.
STEP 4: Draft a follow-up email for each meeting that needs one.
Warm, direct, no filler
Reference specific things from the conversation
Save to drafts only. Do NOT send.
STEP 5: Update your context files.
For each person you met with:
If they are already in your contacts file: update with anything new you
learned — role changes, priorities, what they said they would do.If they are new: add them as a new entry with name, company, role,
and what you know from today's conversation.
Update [your project status or running notes doc] with a dated entry:
what calls happened, what was decided, what to watch for next.
STEP 6: Write a brief output report and show it to me.
Calls processed (meeting name, attendees, key topics)
Tasks created (task title, project, due date)
Email drafts saved (recipient, subject — NOT the content)
Contacts updated (existing records updated + any new people added)
Any signals worth flagging: new leads, relationship changes, things to watch
Needs your attention: unmatched tasks, unrecognized people, anything that
needs a decision before it can be processed
If no meetings were found today, write one line: "No calls recorded." and stop.
Sign off on all emails: [your name]
The context file in Step 0 is what makes this feel like a project manager who knows your work, not a generic assistant. It is a document you write once that tells AI who you are, what you are working on, and who the people in your world are. A Google Doc works. Even a well-organized text file works. Without it, AI treats every meeting like a first encounter. With it, it knows that when a certain stakeholder's name comes up, there is history and a relationship to take into account.
🎓 Coming up: AI Fluency for Senior Leaders
Want to build this live with me? In Week 2 of AI Fluency for Senior Leaders, we do exactly that. One session, your tools, your context files, your workflow running before we log off. Not a demo. Your system set up. As a Clarity Prompts subscriber, you get 30% off.
📡 What Leaders Should Know This Week
Microsoft unveiled Scout at Build 2026, a proactive AI agent that monitors your schedule, clears busywork, and surfaces to-dos through Outlook and Teams without being asked → If you're a Microsoft 365 user, an AI agent is about to start monitoring your calendar and inbox and jumping in without being asked.
Microsoft Copilot Health now pulls your wearable data, lab results, and medical records into one secure hub tied to 50,000+ US providers → If you use Microsoft 365, you can now ask Copilot questions about your own health.
OpenAI gave Codex six role-specific plugins and launched Sites, which turns a plain-English prompt into a deployable, shareable app → You can now build a working dashboard, tracker, or app just by describing what you want and then deploy and share it with your team via a link.
💬 Your Feedback Matters
A big THANKS to those who filled out the survey!!
If you have not responded to the Clarity Prompts survey yet, I would love your input. Two minutes, and it shapes what comes next.
Found this useful? Forward it to one person who needs to hear it. Getting this for the first time? Subscribe free to receive this weekly newsletter here.
↗︎ Think Better. Execute Faster.
See you next week — Purti
Clarity Prompts The weekly AI operating system for sharper leaders.
