Writing on AI workflows and operations

I'm a self-proclaimed fixer. I hear complaints and frustrations about process every single day, and I've learned to treat each one as a signal that there's a gap here worth closing, whether that means more training or a hard look at how the system is set up.

Recently I noticed our executors were manually changing task statuses based on feedback and approvals. I asked whether that step could be automated. First question: is it even worth exploring? If so, what does the solution look like?

Now that approval step is largely automated, and our executors get that time back to focus on the actual work. Even the smallest process change can make a big difference.

See the full post on LinkedIn →

AI accelerates a process, for better or worse. It doesn't fix one. A team that layers automation onto an unstructured intake process just gets faster at being disorganized.

It took me a while to work out how to bring AI into our workflows. What I've found is that if there's a gap, AI will find it and expose it. As AI becomes part of more of our roles, that's the opportunity: use it to surface the process gaps, then fix them.

See the full post on LinkedIn →

Some of our projects run well over a hundred tasks. The people executing those tasks don't always see that a project is a priority. They're focused on what needs to get done today.

I used my Claude Wrike consultant to design a fix. It took two Wrike AI agents plus dozens of automations on custom item types. Agent A watches for new items so each task inherits the project's priority. Agent B does the same for anything added mid-stream. Automations then fire on the task's priority to @ mention the assignee about the priority work they've just picked up.

The one catch: we have dozens of custom item types, so the automation has to be duplicated for each one. Fingers crossed that applying a rule across multiple CITs lands on Wrike's roadmap.

See the full post on LinkedIn →

It's easy to pull a quick dashboard and use the data to confirm something you already believe. Why is my project delayed? Why is my workload heavier than someone else's? Why is that team always at capacity?

At face value the data tells one story. It isn't until you dig into how tasks actually relate, especially where dependencies are involved, that you see what really happened.

I give Claude the goal of helping me understand why something went sideways. It doesn't just pull the first cut of data, it looks further into what actually occurred and suggests improvements. Data-informed decisions make the job easier. You can't argue with facts.

See the full post on LinkedIn →

Picture two teams who can't agree on a process. Neither one is wrong, but each needs different tools. These are exactly the problems I love.

AI has become instrumental here. I now have a third party to think through solutions with, ways to let each team keep the tools they prefer. Right now I'm learning how to use Power Automate with the Wrike API to solve one of these, and I can't wait to bring the solution to the teams.

See the full post on LinkedIn →

Most people meet Claude through one model and stay there. It isn't laziness. Nobody walks you through the dropdown or tells you what each option is for. Once I started matching the model to the task, the same prompts started producing better work in less time and at lower cost.

The rough version I use:

• Haiku when I already know what good looks like and just need volume. Thirty subject lines, alt text for a whole asset library, tagging a backlog.
• Sonnet for the normal week. Status roll-ups, drafts, brief QA, turning meeting notes into tasks. This is my default.
• Opus when getting it wrong costs me something. Board-level documents, architecture decisions, multi-step agent work like pushing 100+ projects into Wrike through our MCP without babysitting it.
• Fable for the genuinely hard problem nothing else has cracked. I don't reach for it often, and that's the point.

Try this: take one task you do every week and run it on Haiku, then Sonnet, then Opus, back to back. Within ten minutes you'll know which one that job actually needed.

See the full post on LinkedIn →

Seeing everything the Claude and Wrike MCP can do has been eye-opening. Last week I built a project with over 100 tasks straight from Claude into Wrike. I handed it an Excel sheet and told Claude how to read it:

• Column A is the task name
• Columns B to F are the description, split into paragraphs (yes, formatting works)
• Link each task back to the parent project for easy navigation
• Add a file reference link to every task
• Batch tasks into scheduled time blocks
• Assign the user
• When it's done, drop a link to each task into the project description, grouped by date batch, so the requestor can navigate everything

Hours of work, done in minutes. I even set up a status dashboard in Claude that tracks project health weekly and sends me a scheduled report.

Start with the executional parts of your role that take the longest, then just ask your AI tool how it can help. Once I got past the initial overwhelm, it's been one idea after another.

See the full post on LinkedIn →

In Wrike we keep regional versions of many blueprints so we can handle pre-assignments. In theory they should be identical apart from who's assigned. Doing some blueprint hygiene, I found three that should have matched but clearly didn't. Some had more tasks than others.

Instead of comparing all three by hand, I dropped them into Claude and asked it to find the differences. A few minutes later I had everything I needed to keep moving. Even something this small can be a huge timesaver.

See the full post on LinkedIn →

PMs hear it constantly: "it's not a heavy lift," usually when someone wants to change scope midway through a project. What requestors don't see is that there may already be hundreds of projects in play, and even a small change ripples into all of them.

This is where communication, prioritization dashboards, and cross-team collaboration matter. I don't like flatly saying no. I'd rather inform and lay out options for next steps. Most people entering a project are only thinking about their own priorities. PMs hold the bird's-eye view and can keep the train on the tracks.

See the full post on LinkedIn →

Is there anything more exciting than teaching a consultant something new? I worked with a Wrike consultant who was excellent at helping build a new framework in Wrike, and one thing she pointed out was my use of the Wrike Database features.

In marketing, specs change constantly, especially for social. I got tired of updating every blueprint each time a size changed, so I built a Creative Specs Database linked to each creative asset custom item type across our blueprints. Now when a spec changes, it's one change in the database and it updates across the entire platform, saving hours of rework.

I love that Wrike keeps rolling out features that let us keep scaling our work.

See the full post on LinkedIn →

I accepted a project without fully understanding how much the scope had shifted. I came back from leave to find it had grown from a handful of projects to over 100, all of which needed to go into Wrike.

Instead of entering everything by hand, I used Claude to clean up the data, and through our Claude MCP it pushed all the projects into Wrike for me automatically. What would have eaten up hours was done in minutes.

It's a great reminder of how AI can accelerate our workflows, especially by automating the small, time-consuming tasks that would otherwise take over your day.

See the full post on LinkedIn →

I keep meeting people who want to use AI but feel stuck at the 1:1 chat. They get the conversation, but going beyond it feels intimidating. A developer might call my approach simplistic, and that's fine. I learn by doing and by not being afraid to make mistakes.

When I don't know how to do something in Claude, I just ask it: "Walk me through step by step how I should do this." If a response doesn't make sense, I ask it to clarify. That's really all it takes when you're starting out and trying to understand how the models work.

To help others get past that first hump, I take my own setups and ask the tool to write markdown files and instructions for my coworkers, so they get a faster start than I did. That's how adoption actually grows. You don't leave each person to figure it out alone. Sometimes you have to break it down so the least technical person on the team can pick it up and run with it.

See the full post on LinkedIn →

With Fable 5 about to move to API, I spent my last stretch with it asking one question: where are my own workflows inefficient? It started building skills based on how I actually work, including one that helps other models think more like Fable and produce better output.

A lot of people with Claude licenses have never even tried Fable. The only real limit is your own imagination. In a short window I built:

• An OFT newsletter builder that exports to Mac and PC
• This portfolio site, something I had never made time for
• A European Portuguese learning app, since every other app teaches Brazilian Portuguese
• A social bookmarking platform from one simple prompt
• A code and security audit of my projects
• The bones of a credit card points platform

Here is how it connects to my day job. Every mistake, misprompt, and bad instruction in a personal project teaches me something I can carry into how I use Claude at Blue Yonder. I have started building AI Advisors for the tools I manage, training Claude to pressure test agentic solutions that actually move the needle. AI is meant to work alongside us, not replace us.

See the full post on LinkedIn →

Everyone wants AI to do the impressive work. Strategy, creative, big decisions. But the best first agents are boring: chasing approvals, compiling status updates, checking that a request form is actually complete.

Those small tasks are where a team's week quietly disappears. They're also the easiest to automate well, because the rules are already clear. Nobody debates what a complete intake form looks like. Someone just has to check.

My 7-agent playbook started there. Intake checks, request routing, deadline alerts, follow-up nudges. None of it demos like magic. All of it gives hours back every week. Start boring. The impressive stuff can wait.

Every week I see a team bolt an AI agent onto a workflow nobody has looked at in years. The result is predictable. The same bad handoffs, just faster. The same unclear ownership, now with a bot in the loop.

At Blue Yonder, the 80% productivity gain didn't come from a tool. It came from mapping how work actually moved, cutting the redundant steps, and only then putting a platform under it. Automation was the last step, not the first.

Before you build an agent, put the workflow on paper. If you can't explain the process to a person, you can't explain it to a machine. Fix the process first. Then automate it.