This was originally published on All Agents Considered.
I’ve spent months teaching Hermes the shape of my work, from where each project lives to the point where it should stop for a decision.
Last week I reversed the arrangement by giving Hermes a limited view of my recent work and asking it to describe the person behind those files.
My official title is full-stack AI engineer, but Hermes came back with something far less suitable for LinkedIn:
AI Workflow Field Reporter With a Side Cart of Websites
The phrase was ridiculous enough to make me laugh, yet it was accurate enough to keep me reading.
In This Edition
Hermes used my recent files to describe what I appear to do all day.
The audit exposed a gap between my engineering title and my visible output.
Editorial work left a clearer trail than many technical decisions.
The read-only boundary hid large parts of my actual working day.
You can reuse my prompt to audit a safe corner of your own workspace.
Here’s How It Works
Most AI tools learn about us through the tidy summaries we write before asking them to do any work.
A profile explains the job we claim to have, while a project note gives the agent a clean version of whatever we’re trying to finish. The result is a useful identity for the agent, although it rarely resembles the mess of an ordinary week.
On a normal day, I can move from researching an AI release to debugging a TypeScript problem without feeling that I’ve changed jobs. By evening, the useful part of either task may have become an article that still needs editing before I can publish it.
I wanted to see what Hermes would say if I stopped describing myself and allowed the artifacts to make the introduction.
This exercise was possible because most of my agent work already lives in plain files that both Hermes and I can inspect. I described that setup in Why My Best Agent Workflow Is Mostly Files, where a small amount of structure made the workspace easier for an agent to follow.
For this audit, I limited Hermes to the parts of my vault that record published work and the planning behind it. I also included working notes and a small set of professional documents, although I did have to sanitize the answer.
I treated each file modification as proof that some work had happened, while leaving effort and time outside the audit. Also, every conclusion needed a visible piece of evidence from inside that boundary.
The Job Title
Hermes sorted my work into a few basic categories and I observed that the actual connections mattered way more than the labels.
Publishing and research created the most visible output. These are the articles and guides people actually read.
Workflow design sat right underneath that surface layer. It mapped out how every project should move and also decided exactly where human input was needed.
The audit also looked way beyond my newsletter (I generalized that outside work for privacy). This gave Hermes the full picture of my career., basically stopping the system from treating AAC like an isolated hobby.
Seeing everything together produced a surprisingly accurate description. I never would have written it for myself.
You appear to be an AI workflow educator and operator who turns technical research and firsthand build stories into practical published guidance, while keeping several professional and commercial projects moving in the background.
That sentence would look totally ridiculous on LinkedIn, but it perfectly describes my actual daily life.
The Strongest Trail
The playful title gave me the opening for this article, but the strongest finding came from the evidence underneath it.
Hermes found the clearest record in my editorial work because I preserved the full path of each article from uncertainty to publication.
I traced ideas taking shape in the early files, then saw which claims actually survived my checks before publication.
Hermes saw way more than finished articles because I had documented the whole system behind them.
Hermes called the pattern editorial systems work , and the phrase stuck with me because people only see the writing at the end. I spend most of my effort deciding whether an idea deserves attention or whether a claim has earned the confidence I want behind it.
Part of my editorial work involves removing weak signals before they ever reached a draft, then shaped the surviving claims into explanations another person could follow.
I expected building to dominate the audit, but Hermes found that I built a lot of infrastructure just to support the workflows around my technical work.
Reading the files back, I saw engineering and writing as different stages of the same loop rather than two unrelated jobs sharing a calendar.
I build until something teaches me a useful lesson, then I preserve that lesson in a form another person can inspect.
That description fits AAC way better than saying I write about AI.
Uncovering the Weak Points
Hermes also found a gap between the identity I lead with and the work that left the most evidence behind. I foreground engineering because I build production AI systems that extend from the interface to the infrastructure underneath it.
But the recent artifacts told a broader story. I kept turning that technical work into explanations another person can use.
Inside the selected folders, engineering showed up more as the thing I was explaining than the thing I was building. That observation stung on the first read because it sounded like a challenge to an identity I know is accurate.
The audit boundary explained most of the tension since editorial work leaves a thicker paper trail than most engineering decisions. I might produce several visible drafts for one article, while the hardest technical decision of the week survives as a single changed line in a repository outside the audit.
Visibility favors work that documents itself, and that gave the editorial side of my week an obvious advantage.
Even with that limitation, the audit made it harder for me to dismiss publishing as a layer wrapped around my real work.
What Disappeared
Keeping the audit safe forced me to hide a huge part of my working day from Hermes. I kept my engineering work in separate folders and repositories and left almost no trace of my private interactions. This blind spot also swallowed any experiment that failed before I wrote down the results. Hermes just described my best-documented work instead of reconstructing my actual calendar.
The system captured that limitation perfectly in one line:
A file trail records exhaust. rather than the engine.
That distinction keeps the exercise useful since the agent only describes visible work instead of pretending to value everything outside the frame. A polished plan might represent a project I never started. Meanwhile, a single modest code diff might represent the hardest afternoon of my month.
The audit also exposed a smaller operational problem.
Some article projects contained several drafts without a clear current version. Those files proved I did the editing, but they forced Hermes to waste extra effort deciding which draft actually mattered.
I previously found the larger version of this mess when I ran the twenty-minute audit that showed where Hermes was getting lost. Seeing this exact problem pop up as a footnote today tells me even my digital biographer needs stricter file naming.
There’s A Valuable Lesson Here
The audit gave me a much clearer description of the loop I repeat across all my projects. A technical problem always starts the cycle, and the useful lesson becomes the only thing worth preserving once I finish the immediate work. I often turn that lesson into an article because writing forces me to explain exactly what happened without relying on the original conversation.
This exact loop explains why I ditched isolated prompts and started building workflows that leave evidence between steps.
I described this process in How to Build AI Workflows When You’re Tired of Optimizing Prompts by showing how one step writes something the next step can read. Hermes reconstructed my recent work entirely from the residue of those handoffs.
That happy accident turned the whole audit into a live test of my system.
The weak spots in the final answer showed exactly where my ongoing projects still need a better handle for the next session. Dropping a small current-state note into the folder usually saves a project when it needs to survive beyond the original conversation. Passing thoughts can still disappear naturally without forcing me to turn my whole workspace into a diary.
And those existing files gave Hermes just enough evidence to spot my patterns (to a point) without relying on constant activity tracking or screen recording.
Your Turn
You can run a smaller version of the same exercise over a few folders you already control.
Choose a recent time window and keep the task read-only so the audit can’t reorganize the evidence while inspecting it.
Exclude private material before the run, then require the agent to cite the files behind each conclusion.
This is the prompt I used:
I want you to describe what I appear to do for work based on the
artifacts I leave behind.
This is a read-only audit. Don't create, edit, move, rename, or
delete anything.
Inspect work produced or modified during the last 30 days inside
these folders only:
[LIST SAFE FOLDERS]
Skip credentials and private material.
Treat file changes as evidence that work happened, without using
them to estimate time or effort.
Give me a one-sentence description of what I appear to do.
Give me a playful job title based only on the evidence.
Describe the strongest pattern you found and cite the files that
support it.
Explain which part of my work is likely invisible to this audit.
Compare the evidence with my official job title only after reaching
your conclusion.
Separate direct evidence from inference.
End with a short section called "Where I Might Be Wrong."
Keep the tone observant and lightly humorous rather than judgmental.
Read the result as a mirror whose frame you chose rather than a performance review with any authority over your working life.
Correct the parts it gets wrong before removing anything too private to share outside the workspace.
Then pay attention to the work it can’t see, since missing evidence may explain more than the confident conclusions.
Behind the Title
My official title still says full-stack AI engineer, but the weird title Hermes generated remains way too strange for my next family gathering.
That strange title highlighted a habit I usually downplay because my writing grows naturally straight from my technical work.
My commercial projects sit right next to those two main areas since I apply that exact same habit to my business work. I always turn useful builds into plain explanations that other people can actually use.
My public work creates very clean file folders while private conversations leave almost zero evidence behind for an AI to read. Hermes only analyzed this heavily documented version of my career, yet I still completely recognized myself inside that incomplete picture.
This exercise proved I built a clear enough trail for a machine to follow without pretending this final summary captures my entire reality. My official job title just names the foundation, and these files show exactly what I keep building on top of it.
Top comments (0)