ProductivityPrompt - Which mail should I read!?

#genAI is everywhere on LinkedIn. How much of it is already in your daily lives and routines?

I've been using #LLMs for many months now (and related technologies for years). I love the trend towards more local usage of them. So I've started to challenge myself to use more #localLLMs in my routines. And to share these experiences with you.

The use case for the ProductivityPrompt

One of my biggest challenges every day is slicing my time-in-a-day and focus on the right things. I receive MANY mails every day and I asked my model to summarize, prioritize and extract the ToDo list for me.

I'm fairly happy with the results after some experiments using Llama 3.1 8B.

My technical setup

I'm using llama.cpp to run my model as it is flexible enough to use GPU or CPU depending on the available memory.

I'm using Meta-Llama-3.1-8B-Instruct.gguf as a model as it offers up to 128k tokens context size (so I can just drop my mails) and it's of fairly high quality. I would have loved to use a bigger model but my laptop only 20-odd GB of RAM and no dedicated GPU.

The prompt I used

The key ingredient is the prompt as it guides the model on what it should be doing.

For today I used a very simple prompt and Prompt Engineering gurus can point out improvements - I'm happy to update this article if you have great advice!

I need your help to manage my email inbox.
I will provide you with a few email messages and I need you to provide the following:

1. High level summary of each message in 3-4 bullets
2. Main todos for me, Fabian
3. Ranked order of the mails

The ranking is based on the Eisenhower principle.

Here is the list of messages:

** Start of message - [Subject line]**
[message content just copy&paste]
** End of message - [Subject line]**

** Start of message - [Subject line]**
[message content just copy&paste]
** End of message - [Subject line]**

** Start of message - [Subject line]**
[message content just copy&paste]
** End of message - [Subject line]**        

The results

Here's what I got for three random mails in my inbox.

Now it's time for you (Fabian) to take action:

1. High level summary of each email in 3-4 bullets
2. Main todos for Fabian
3. Ranked order of the emails based on the Eisenhower principle.

Please provide your feedback, and I will proceed with ranking the list accordingly.
        

The model repeats the mission statement and even invites for an iteration. Nice touch!

### Email Summary (High Level)

**Email 1: SAP [redacted]**
* SAP TechEd registration is open
* OpenSAP platform [redacted]
* Capgemini and SAP announced a joint press release about AI
* Reminder for [redacted] RISE 

**Email 2: Daily Cadence- [redacted]**
* File shared with roles that need to be populated by the team
* Roles can have seniority levels like discussed in today's call
* Clarification needed if any further details are required

**Email 3: RE: [redacted] awaiting your approval**
* Request for exception to simplify [redacted]
* Challenge of [redacted]
* Requesting support and looping in [redacted]        

It nicely picks up the names of the mails and lists good summaries. I've redacted some content for privacy reasons but I think you get the idea of the level of details and the style.

Now to the actionable part:

### Main Todos for Fabian

**Email 1: SAP [redacted]**

* Register for SAP TechEd if interested
* Review RISE [redacted] 

**Email 2: Daily Cadence- [redacted]**
* Populate roles in the shared file as per requirements

**Email 3: RE: [redacted] awaiting your approval**

* Discuss exception with [redacted] to simplify [redacted]
* Review [redacted]        

The model was a bit too eager to find ANY kind of to-do in the first mail (which had no real to-do for me) but overall it was correct in the actions it suggested.

An now the most important part for me - where should I put time and effort? I cheated a bit by using Eisenhower on a set of mails without indication an what COULD be delegated. So the model predicted the best it could:

### Ranked Order of Emails (Eisenhower Principle)

**Urgent & Important: Email 2: Daily Cadence- [redacted]**
Population of roles is required, and a clear deadline might be involved.

**Important but Not Urgent: Email 3: RE: [redacted] awaiting your approval**
Discussion with [redacted] requires attention to resolve the challenge.

**Not Important or Urgent: Email 1: SAP [redacted]**        

Conclusion

I'll mark this one as a net win for me in terms of productivity. I'll need to run a few more tests but I strongly believe this to be helpful already. After some more prompt engineering and by injecting some context about the things I care about, the LLM should be fairly effective in predicting the order of things.

BTW - my Outlook is configured in a way that messages from certain people or certain organizations always show up in red in the message overview. While I trust the LLM to do a good job, I will still provide a bypass for some people!

#PeopleMatterResultsCount #LifeAtCapgemini #ProductivityPrompt