For about 10 years, I have been presenting a CLE on what lawyers should and should not do when responding to negative online reviews. The ethics opinions and case law in this area keep moving, so the webinar needs regular updates. This week I rebuilt my presentation with Claude, and the workflow is worth sharing. I have written before about how I use AI to create PowerPoint decks, but I wrote that post several months ago. As AI has improved, so has my process.
Step One: Instructions
It is important to give the AI instructions about how to craft the PowerPoint. I give the AI standing orders for deck building so the tool will know what I want. For example, I keep my font size at 20 or larger for most things. This information can be placed in standing orders (if the tool has that ability), in the instructions for the specific project, or in the first prompt. This is also where you tell the AI about your branding. In many cases, you can just point the AI to your website, and it can copy your branding from there.
Step Two: Creating a Project
When I create a PowerPoint deck with AI, I start by creating a project. I normally use Claude, but many AI tools now have a project feature. Google’s NoteBookLM also is a great tool for creating slides. The key is the ability to give the AI access to all of the sources in one place so the tool can access everything we need. I will note that I rarely use Copilot to create my PowerPoints. Copilot lacks the flexibility of other AI tools for this purpose.
In the case of my Bad Review deck, I started by creating a project in Claude Cowork. You can use Projects in the web-based version of Claude or the desktop version. The desktop version includes Cowork so that is what I used.
Projects let me load reference materials once and reuse them across every chat in that project. Cowork lets me point Claude to appropriate folders, so I don’t need to physically upload the documents myself. For this presentation, I pointed Cowork to my PowerPoint folder and also a folder where I keep relevant ethics opinions. For folders I did not want Claude to access wholesale, I manually uploaded only the specific files I wanted it to see. The data we began with included the old PowerPoint deck, notes about improvements I wanted to make, relevant ethics opinions, articles I had written on the topic, and a stack of recent cases.
Step Three: Creating and Editing an Outline
Next, I asked Claude for an updated outline based on everything to which I gave it access or uploaded to the project. The first draft was solid but not quite right. I told Claude where I wanted to change emphasis, what to cut, and what to expand. We went through the outline together until it matched what I actually wanted to teach. This is the collaborative process necessary with AI. I don’t just accept its output; I work with that output until it is accurate and says what I want it to say.
Step Four: Identifying Additional Sources
Before letting Claude write anything more, I asked it to list every source it had drawn from or planned to draw from. Claude pulled together a list that included the materials I had provided plus several others it had flagged as worth checking. We talked through which sources I actually needed, and I read each one myself to confirm accuracy. A few of Claude’s suggested sources turned out to be tangential and we dropped them. Others were squarely on point and went in. I added all of the sources I decided to use into the project so Claude would have access to them as we worked together. Giving Claude the documents helps reduce the risk of hallucinations.
Step Five: Creating the PowerPoint
Once the outline and sources were settled, I told Claude where I wanted to make additional changes, reviewed the revised outline one more time, and then had Claude build the PowerPoint.
Then I did something I have started doing on every significant project that is not confidential. I took the outline to both Grok and ChatGPT and asked each one what was missing or weak. Some of their suggestions were genuinely useful. Others were the kind of generic advice that sounds good in the abstract and adds nothing in practice. I brought every suggestion back to Claude and we went through them one by one. We talked about the benefit of adding each idea, kept the ones that fit, and dismissed the rest. Again, this is how I collaborate with AI. I don’t accept its work without critique, but I accept its advice when I think appropriate.
Claude rebuilt the deck. I reviewed each slide and told Claude what I wanted changed. It is important to understand that AIs will keep working and giving advice as long as I keep talking. So, with each change I gave Claude the instruction and then told it to wait until I was ready before doing anything else. That kept me in control of the pace and stopped Claude from running ahead and rewriting things I had not finished reviewing. If you don’t tell Claude to hold until you are finished, it will keep recreating the PowerPoint, which will eat up all of your tokens.
The morning’s work used up my tokens by around 11 a.m., and I had to stop until they refreshed. Once my tokens refreshed at around 2:30 p.m., I resumed my work. The deck got built in a fraction of the time it would have taken me alone, but the heavy back and forth burns through usage quickly.
Suggestions For Those Who Want to Follow My Process
A few things I would tell anyone trying this workflow.
- Load everything into the project before you start. Adding documents mid-conversation slows Claude down and makes the output less consistent.
- Ask for the source list early. If you do not check the citations yourself, the whole exercise is worthless. Claude can hallucinate sources, and so can the other models. I read every opinion and case before I trusted it.
- Use more than one model for review. Grok and ChatGPT each caught things Claude missed, and Claude caught things they missed. They are not interchangeable.
- Make the AI wait for you. The default rhythm with these tools is to keep generating. You have to slow them down.
The final deck is an improvement on the last one. The AIs recommended issues I might not have thought of and surfaced resources I might not have considered. Since I read all sources, I learned more than I knew before. My work with AI was, as I noted, collaborative in nature. That collaboration provided me with new ideas, new material, and saved time. But I was in charge of the collaboration, and nothing gets shared until I am certain in the factual claims and the appropriateness of the sources that I used.
One last thing, if you are interested in my webinar, I offer it through https://clewebinars.com. You can sign up to take it on their site.