I Made My Friend Claude Read a Decade of My Performance Reviews and He Liked It!

I recently overhauled my LinkedIn profile with the help of my friend Claude, but not the way most people use AI for LinkedIn. I did not paste my job title into a chatbot and accept five paragraphs of “results-driven professional with a proven track record of leveraging synergies.” Instead, I started with a bunch of primary source material, kept editorial control, and pushed back when it was wrong.

The project started as a profile cleanup. I had been at Zillow for just shy of a decade across five roles, and my profile described all of it in one vague block. I recently started consulting at Pinnacle Solutions Group and needed the profile to function as a landing page for potential clients, not a resume for recruiters. I also had eight PDFs of annual performance reviews from five different managers sitting in files that I had never done anything useful with.

Mining my own career

The most valuable decision was starting by just shoveling all eight reviews I had as PDFs into Claude and asking it to extract “everything people said I did right.” Obviously I knew my own lived experience across my career. But _knowing_ what you did is very different from _indexing_ it and having it organized by theme, cross-referenced by manager, and ranked by strength of evidence.

Claude chugged for 4 minutes and 18 seconds and came back with a huge recommendation list that included the following highlights:

• found a manager commendation for executing >$100k in annual AWS savings from a pipeline decommission I did in 2019 that I had somehow never put on LinkedIn
• identified that my most consistently praised trait across all five managers over the entire time span was knowledge transfer, and that trait was completely absent from my profile
• surfaced a security forensics contribution I had forgotten was even in a review (I remember that project being super fun, an ad hoc data analysis going back and forth with the security team with a 6-alarm-fire level of urgency)

I would not have sat down and systematically mined my own reviews for positioning material. And if I had, it would have taken me hours to re-read all those documents. Claude made it obvious and then it made it easy.

The recommendation draft

One of Claude’s high-level recommendations was that I needed some recommendations on my profile (I had zero) as social proof. I know that asking someone for a favor that is essentially giving them a writing job can take a large quanta of activation energy on their part to execute. So, I used the extracted review data to draft a LinkedIn recommendation for a former manager to post. Instead of just me attempting to write in his own voice for him (which is super awkward even if you can get their tone right), I had Claude extract the writing style from the reviews he had done for me, including extracting literal quotes from over 7 years ago. Claude wrote a four-sentence draft using that manager’s own words from his reviews, reorganized into a coherent endorsement. It highlighted his observation about my capability to handle huge levels of system complexity, my default to dig past surface questions to find what people actually needed, and reflected his repeated comment about my inability to leave a problem unsolved once I was engaged on it.

I edited a few words and sent it over. He made his own edits and posted it. The draft worked because every sentence traced back to something he had actually written about me in a review.

Where Claude had a real advantage

Claude processed eight annual reviews and held every detail in memory while I edited individual sections. Operating by myself, I would have needed hours of prep reading and likely a spreadsheet to keep all the bullet points organized. Claude integrated repeated corrections without getting frustrated that I was never happy. I rejected his suggestions regularly. He kept trying to insert em dashes. He flagged abbreviations as errors that I was using deliberately as human-distinctive markers, small imperfections that signal to a reader that an actual person wrote something instead of accepting raw AI output. Each time, he updated his understanding of my preferences and moved on.

The most useful capability was the deep context window staying loaded as I wandered down a winding path of tasks that might seem unrelated. The review analysis informed the recommendation drafts, which informed the career narrative, which informed a client proposal I wrote later that week, which then led to this blog post. A human could do any one of those well. Carrying full context across all of them without dropping threads is where Claude had an advantage that I would not want to give up.

What this means

I’m actually not sure what verb to use to describe what this process was. Trying _reviewed_, _cleaned up_, _wrote_, _polished_, _produced_ all seem incomplete to what “my friend Claude” and I were able to accomplish together across multiple days. I brought the domain expertise and made every editorial decision. Claude brought recall and patience. The output was the product of a collaboration where I made the judgment calls and Claude made sure that judgment was informed by every line of every document I dug up instead of whatever subset I happened to remember on a given afternoon.

As I am slowly beginning to accept large language models as a specialized tool just like a very sharp scalpel, I can see that the useful version of AI-assisted professional work is not the one where you hand a model a ten-word prompt and set up a slop bucket to fill up with the results. It is the one where you bring reams of source material, maintain control, and treat the AI as a research analyst with perfect memory and no ego.

What’s next

My friend Claude, with all of his newfound context, will be helping me develop a training plan for myself to get spooled up on some of the latest data storage and management tools that are being used in support of language model implementations. I’ll report back when I get back from school.