Key Takeaways

  • Daniel Blum, product manager at Melio, runs an automated morning brief using Claude and Cowork that synthesizes Granola transcripts, Slack messages, and emails into one-line summaries with clear action items.
  • When the model encounters company-specific terms or unknown project files, it flags them and proactively asks Blum for definitions.
  • Confirming a definition once (such as a payments term like "settlement cap") stores the concept directly into persistent context files for future tasks.
  • Claire Vo points out that standard foundation models fail at work because internal company lingo does not exist in base training data.

The Method

Most teams feed context into AI tools backwards. They write static documentation, paste it into custom instructions, and watch it go stale within weeks. Blum flips this workflow around by turning morning triage into an automated context-gathering engine.

Every morning, Blum runs an automated daily brief across his primary communication channels. The workflow ingests three data streams: audio transcripts from Granola, unread threads in Slack, and email updates. Instead of just summarizing what happened yesterday, the system runs an active audit for things it does not recognize.

As Blum explains:

“In this daily brief, my coworker runs through my recent slacks, my emails, my personal notes, and it looks for context that it doesn't understand in these things. And it says like, oh, if there's a file I wasn't aware of, there's a milestone, there's a goal, it would ask me about it and say, hey, like, what is this? Is this important? Should I read this?”

During one morning review, the system caught a payments term it had never seen before: "settlement cap." Rather than hallucinating a guess or ignoring the thread, Claude extracted what it could from surrounding messages and asked Blum directly:

“Here it flagged something. Let's see what it is. A term I didn't know the settlement cap. All right. This is some payments thing and it says like I read the thread and understand it's a limit working through. Want me to save? Yeah. So I'll say save. Yes. Save it to the context.”

Vo highlights why this matters:

“And so many of us work inside companies where the things that we say are not in the training data. They well not yet at least they're not in the training data. And so actually proactively prompting Claude to say like I don't understand what you're talking about here. Can we define it together and then I'll save it to my context and I'll know what you're talking about moving forward is really really sharp”

By treating the model as an inquisitive assistant that actively builds its own dictionary, Blum avoids manual prompt engineering. The context file grows more accurate every single day.

Where This Breaks Down

This method requires clean raw inputs and clear boundaries. If your Slack workspace produces hundreds of noisy, automated bot alerts every hour, your morning brief will choke on irrelevant text strings and trigger false alarms.

Second, persistent memory without validation creates risk. If you quickly click "save" on a rough definition while groggy at 8:00 AM, the model treats that mistaken definition as ground truth for every downstream PRD, email draft, and summary. You must treat context confirmations as permanent configuration changes.

What to Do With This

Open your primary LLM workspace and add an explicit rule to your daily summarization prompt: "Identify any internal acronym, file name, metric, or project codename in these inputs that does not appear in your persistent memory file. List them at the bottom of the brief with your best guess from context, and ask for my confirmation to save them." Run this on tomorrow's first meeting transcript.