We have thousands of images in a bunch of subfolders with random names. How do we figure out which one will be useful for a social media post?

This was the question that kicked off a discussion with a small business owner. They had spent hours working on a strategic plan and social media strategy, and they had thousands of images from events over the past few years in their Google Drive. However, they had not named the photos in a way that allowed them to easily figure out which one could go with a post.

Workflow of a photo processing pipeline using claude cowork

As they described this business problem, I suggested they make use of Claude Cowork to tag these photos. Since they had already written the social media posts, we weren't going to use this generative AI tool to create any AI slop, but to help them automate a business problem, large number of photos in an unstructured file format.

Prompt: "I am creating this project to help with social media posts. I want to develop 20-25 one word tags to help me identify photographs. Ask me questions to come up with the tags. They should be a mix of who's in the photo (group, solo), orientation (portrait, landscape), content, etc"

We started using Claude to interview them about what they were looking to capture in each image:

  • orientation (portrait versus landscape)

  • groups versus solo photos

  • types of artwork

  • other specific items that they would find in each photo

CSV output of the photo tags used in the pipeline

The output captured as a durable CSV file which could be passed to the next request in Claude Cowork. CSV files are a great way to preserve context between steps of an AI workflow without having to scroll back though chat messages.

Prompt: "Review all the photos in this folder. Create a CSV file with five columns the filename and full file path, a description that I can use for alt text, 3-5 of the tags from the Photo Tags file, and a status (default UNUSED), and event date"

we were then able to move on to the next step, where we directed Claude to review all of the images in a folder. We downloaded the images to the folder. Even though the images were saved locally to their computer, it's important to know that images did have to be sent to servers controlled by Anthropic in order for a large language model like Opus to process the images. These images were captured at events with the intention of their use in social media images, so there was limited security risk of using a remote large language model.

The CSV file of photo descriptions, tags, and image dates

Their request was not for creative work from the model, only to describe what was in each image. Again, we requested a durable output that we could pass on to the next step. This included:

  • the full file path, since there were many subfolders involved

  • the tags that we developed from the previous step

  • the alt text

  • description of the image

  • the date that the image was created

Photo workflow highlighting the pre existing social media posts.

The image metadata was saved into another durable CSV file. The next step in the workflow used the existing social media posts that had been written by their social media staff.

Prompt: review all of the social media posts in the social-media-posts.csv file. Suggest three photos that could be used for each social media post, based on the photo-descriptions.csv file. For each row of the social media CSV file, add three additional columns using the full file path including folder names for each suggested image
The photo workflows final output of selected media posts, and the owners feedback

The requested output from this stage was a pairing of one social media post with one to three images from the described section, again captured as a durable CSV. The owner reviewed this final output to choose, based off of their own preference, which image would go with each social media post.

At the end of the process, not only was there a list of social media posts along with a specific image to go with each, but also feedback about why they picked the specific image to go with each post, and why the other two were discarded. This allows for further improvement of the tagging and description strategy to make sure we can hone in on exactly what the business owner is looking to get at. This repeatable process has allowed their social media staff to focus on posting and less on photo administration.

Feedback from the business owner was that this workflow unlocked their social media posting strategy and saved many-many hours that they didn't have to manually process image after image. AI can help you get work done, but you need to be the one guiding it.