A question I keep being asked in interviews and applications is how I use AI in my daily workflows.
As a software engineer between jobs who never really took up side projects, I have been framing my answers around past employment. But that gap is growing, and AI is advancing.
Sure, I have been running search queries, product comparisons, and financial scenarios through AI these last few months. I know to take the output with a grain of salt, check the sources it gives me, and do my own follow-up research. But that is still closer to natural-language search with a wider net than the problem-solving or technical work I would use AI for at work.
The first context-box surprise
Since Fall 2025, I have used Perplexity for most personal AI use -- s/o to PayPal for the free year trial.
I paused my separate ChatGPT subscription and used Perplexity as a single place for research and conversation, while still being able to choose among models. That arrangement worked well enough for personal use until I decided this past week to start a side project -- gasp! -- and bought a Claude subscription.
When I first opened Claude, it had an example Project already set up: "How to use Claude."
I started a conversation in that Project to get help with prompt design. I began small, with something like figuring out dinner that night.
Very quickly, though, I stopped thinking about prompt design and started thinking about cooking. I listed what I had at home, stepped away from my laptop, grabbed my phone, and went to scan the fridge.
Then I accidentally opened a new Claude chat outside the Project.
I added a few things to the list. Claude told me there was no list.
That was not what I expected based on my experience with Perplexity.
So I did what any reasonable person would do: I asked Claude.
Claude's support documentation explains the behavior: chats outside Projects can be searched together, but each Project has its own separate memory space and summary. Context inside a Project remains focused within that Project, separate from non-Project chats. Claude explains the boundaries here.
That makes perfect sense for business use. A Project per client or workstream is a useful guardrail against sensitive context spilling into unrelated work.
For personal use, I found the distinction jarring. I was not trying to keep one client's information away from another. I was trying to make dinner with the grocery list I had started five minutes earlier.
The kind of memory I wanted
That fridge moment sent me into a larger question: what kind of context should an AI assistant carry between conversations, and when should it be contained?
I resubscribed to ChatGPT. My Projects and memory from last year were still there, but they needed an update. A lot has happened -- and been explained to AI -- since I moved most of my personal use to Perplexity.
ChatGPT's approach aligned more closely with what I wanted. Its documentation describes project-only memory as an option selected when creating a Project, rather than the default behavior. With default memory, non-Enterprise ChatGPT accounts can use saved memories and reference conversations outside the Project, unless another Project is configured as project-only. OpenAI documents the distinction here.
I do not think either approach is universally better.
Claude's model makes clear boundaries the default. ChatGPT's model lets personal context travel unless I explicitly choose to contain it.
For client work, medical information, or a sensitive one-off task, I can see why I would want the first model. For a personal AI assistant, I currently prefer the second.
The side project became a prompt project
That briefly put my side project on pause and sent me down a prompt-engineering rabbit hole.
Inside Perplexity, using GPT-5.6 Terra, I opened a Space for Job Hunting, a high-priority theme for me. Conversations in that Space include resume tweaking, coding and behavioral interview prep, application responses, and figuring out immediate and long-term career goals.
There are plenty of useful details in those chats that ChatGPT did not yet "know."
So I created a new conversation inside the Job Hunting Space and began constructing a prompt that would review the available context. I wanted it to produce a detailed profile and analysis: my job search, work history, current priorities, corrections to earlier assumptions, boundaries around sensitive information, and how I prefer to collaborate with AI.
Then I exported that profile as a PDF.
From there, I wrote a second prompt for ChatGPT. I attached the PDF inside the appropriate Project, told it the document was the current source of truth, and asked it to create or update durable context from it. If an older memory conflicted with the PDF, the newer document won.
Is this just another context box?
Writing that out, I started second-guessing the entire thing.
If I am exporting by Space and importing by Project, am I just recreating the same isolated context that Claude gave me?
Maybe. But I do think some context should travel.
Knowing my current job situation and desired base salary could help a Finances Project figure out future taxes. It could matter in a Mental Health Project where I am reflecting on this period of unemployment and its impact. At the same time, not every detail belongs everywhere, and not every stored memory should become permanent fact.
That is the part I find interesting: the work is not only "give the model more context." It is deciding which context is durable, which is project-specific, what is sensitive, what has changed, what needs verification, and which source should win when two things conflict.
For now, it is also giving me something to do during some humid, cloudy days in July.
What I have built so far
So far, I have:
- Written a detailed, structured Markdown prompt for Perplexity to generate a profile PDF from a Space, using a small set of variables
- Generated the PDF of my Job Hunting profile analysis
- Written another structured prompt for ChatGPT proper to pair with that PDF, import the relevant details, and create or update a profile about me
- Started thinking more deliberately about what should count as durable context versus project-specific context
The prompts currently live in code snippets and still need tweaking. I also need to decide the best way to share them, or whether to share them at all.
I expect this may be a one-time exercise. But who knows? Maybe a new kid on the block will eventually overtake ChatGPT, and I will want a way to move my context again.
For now, this post is a way to document the experiment publicly and show some enthusiasm for this next frontier of tech.
Documentation still matters
Engineering is no longer only coding and problem-solving.
I have always been a strong advocate for documentation and clear communication. That now feels more relevant, not less.
People often say the hardest part of a project is the last 20 percent. I think that remains true, even if the first 80 percent is changing shape. Boilerplate code, familiar patterns, and first-pass implementation may increasingly start as prompts.
The difficult part is still the difficult part: defining the actual problem, identifying missing context, catching a confident mistake, evaluating trade-offs, documenting decisions, and knowing when not to trust the tool.
For now, I have a side project that became a prompt project, a small collection of portable-context prompts living in code snippets, and something to do during humid, cloudy days in July.
I do not know whether I will share those prompts yet. I also do not know whether this will be a one-time exercise or the beginning of a habit every time a new AI product enters the picture.
But I am enjoying the experiment. And, yes, I am a little enthusiastic about this next frontier of tech.
That still sounds like engineering to me.
Top comments (0)