Harleen Chatha is a senior design manager at Canvs, who leads one of India's largest BFSI projects. Over the past few months, she has built a fully agentic workflow that runs round-the-clock in the background. It pulls context, logs decisions, and updates documents automatically every day, and delivers a digest before I even arrive at my desk in the morning. In this piece, she will take you through how she set it up and how it’s proving to be useful for her work.
What does an agentic setup actually run for you day to day?
I have a setup that runs three times a day.
At 9:30 AM: The morning briefing fires automatically. My agents comb through the previous day’s meetings, email, Figma, Slack, and calendar, and I read it on my phone. By the time I sit down at my desk, I already know what is open, what changed overnight, and what needs my attention that day. Plus, all of it is categorised by relevance. This changes the nature of preparedness, before the day starts, before any meeting.
Daily morning Slack digest fired by Claude, with action items, organised by priority.
In the afternoon: There is a check-in that requires me to respond. I confirm what is done, what is still open, what has been dropped. That is the only point in the day where the agent needs something from me.
At 7 PM: An evening digest runs regardless of whether I was at my desk. It ingests all Granola transcripts from that day, then goes through the full CLI route: Google Workspace, Figma, Slack, calendar.
End-of-day-digest which gives a run down of everything that happened that day, plus what needs to be ready the next day.
The output is three live markdown documents that get updated every evening: a daily digest, a decisions log, and a tasks list. They are ready the next morning without me touching them.
Before this setup, I would manually go through everything. The information load was significant enough that it contributed to burnout at certain points.
How much of your context, documented and undocumented, can the agent actually work with?
The documented parts were the easier half. I fed 3.5 years of Notion and MindNode meeting notes from the project into Claude, so all of that (meeting notes, email threads, and so on) is accessible to the agent now.
For the undocumented parts, I run Granola on every call, including casual internal huddles. Everything gets transcribed, and Claude ingests those transcripts every evening. A decision that used to be only verbal is now sourced back to the specific call it came from.
Running Granola notes every call and internal huddle to keep track of all conversations, plus it makes up material for my daily digest's action points.
One example: a particular card product was originally meant to sit near the top of the home screen on the banking app, where it would be one of the first things a user sees. It ended up placed further down the screen instead, below the section a user sees without scrolling, because its sales figures didn't justify the more prominent spot. No one would know that reasoning unless they were in that meeting. Now it's in the decisions log (maintained by the agent), sourced from the call it came from.
Documenting decisions is one of the most crucial parts of my role. Now I have my agentic setup to comb through my emails, calls, slack to extract decision points and log it for me.
Can the agent flag things on its own the way you'd catch them by watching closely, and can it draw on how things played out before?
Partially. It's not fully autonomous, but it's not purely a human thing anymore either.
Within a module, Claude now flags quantitative signals I'd otherwise have to be actively watching for. It flags when too many comments pile up on a certain module of a project which prompted me to check it. Previously, catching that kind of build-up meant I had to be the one monitoring comment counts across every module myself. Now the agent watches for the pattern and only pings me when the number crosses a threshold worth looking at.
What it can't do yet is draw historical parallels. It can't tell me that the last time something similar happened on a different module, a specific outcome followed. That needs history it doesn't have. It only knows what's been fed into it, and that's project-specific, going back three and a half years at most, not further.
So, for me, it’s a half and half.
What was one specific task from the last month where an agent cut your time significantly?
The clearest example is reconstructing context after a handover gap.
A couple of months ago, a colleague left the project. Two months later, I was added to an email thread asking for clarity on a discussion I was not in the loop for.
Before Claude, that reconstruction took four hours and three conversations with another design manager, and it was still incomplete. After Claude, I pulled everything tagged with the relevant shortforms across email, Slack, and meeting notes in under a minute.
Does the agent adapt to your specific way of working?
It's calibrated entirely to how I speak and how I work. I spent significant time upfront training Claude on my communication style. The documents it produces have to meet that same standard. If the output is full of filler, the setup has failed. I also have a hard rule: Claude never responds on my behalf. It can suggest responses, but I approve everything that goes out.
I have input details about me into Claude's memory, so that everything it does is aligned with the way I work.
At some point during setup, Claude described my communication style back to me, and my reaction was something like, "I don't think anybody has understood me like this."
What would you tell a design manager when they ask you "where do I even start"?
Start with Claude's Cowork. I spent two weeks on the free tier before committing to the $20 plan. Give yourself a couple of days to read up on agentic workflows. It can be on Twitter, Reddit, wherever people are documenting their workflows. Then get on the paid plan and split your usage into 2 halves of the day to avoid hitting limits.
The first question to answer is where your work actually lives. Where are decisions being made? For me, it was Slack, Google Workspace, and Figma. The MCP connectors were too token-heavy for my usage pattern, so CLI was the practical path.
A summary of my agentic setup which does most of my admin heavy-lifting work.
The second question is what you want to get out of it. My 2 anchors were tasks and decisions. Everything else followed from those.
I'd also tell them to ignore the generative and visual output for now. It's not relevant to a design manager's day-to-day work. Save it for evenings and weekends if you're curious. The foundational automation is what matters first.
What do you, as a DM, actually do with all the time that frees up?
Three things, in my case.
It gave me more bandwidth for actually checking in with people. For a long time there was no room for basic human conversations, casual catch-ups, just seeing how someone was doing, because the monitoring, the context-switching, and the prep were consuming all of it.
The second is going back to design. I had not made a product myself in months. I am nothing if not a designer. I only exist here because I went through design school. I will not forget that. I have to keep going back to that.
I can fully invest in the parts of this role that actually require a design manager. The parts that require judgment, business understanding, and client relationships are exactly what got deprioritised when project management was consuming everything. I do not need to do project management anymore. Agents are doing that for me. I have to take 2 minutes out of my life 3 times a day to read it. For me specifically, that has meant going deeper into the business of design: understanding the commercial reasoning behind client decisions, the strategic layer that sits above the screens.
Key takeaways:
The pipeline runs three times a day, and only one of those runs needs anything from me. Morning and evening are read-only. The afternoon check-in is the only point where the agent asks me something back.
The documented half of my context was solvable by feeding in history. The undocumented half required a change in behaviour going forward. Three and a half years of meeting notes closed one gap. Running Granola on every call, including the small internal ones, is closing the other.
Flagging a threshold is easy. Recognising "this looks like something from three years ago" is not. That's still on me.
Automation freed up time for the parts of my role that were getting squeezed. That's people, design, and the judgment-heavy work that project management overhead had been displacing.
