How your AI keeps Forgetting (and, how I fixed it)

Hi, there!
Welcome to the 3rd edition of Work in Beta.

In this edition, we are going deep in one specific topic - Projects.

One of the most powerful (repeatable system rooted deep in context) but completely under-used feature of your AI.

So, let’s dive in!

Till early 2025, I had this conversation with ChatGPT at least 50 times:

“Remember, I am a Product Leader. When I ask for deep research, structure it as a product strategy. Format: problem-solution-reference. Keep it under 5 pages.”

Except it didn’t remember this even a single time.

Two days later, I was back at reminding the same thing. Again.

It took me a while but I eventually realized that it is a setup problem.

People treat AI as a fancy Google. They start a new chat everytime, explain their context all over again in a repeat loop. That is massive under-utilization of the tool.

Quality work with AI needs the following:

Here’s what actually fixes the problem: Projects.

Not better prompts. Projects are persistent workspaces where the AI remembers your instructions, has read your files, and doesn’t make you start from scratch every single time.

Let me show you how this works.

A Project in ChatGPT or Claude has three persistent layers:

You need to set up and manage all the three layers properly to truly unlock its benefits. Most people who use projects don’t add any instructions or files, or they add vague or outdated details.

Then they wonder why does the output sucks.

Here’s the actual architecture that works from my tried and tested system perfected over the last 12 months.

I use a lot of prompts for my regular work that take a considerable time for me to put together manually. I need to add context, instruction, templates that remain constant across the prompts I build. Copy pasting is a lot of time wasted and duplication of effort. I could do this inside a single chat, but it starts hallucinating as the context keeps growing.

So, after going through months of pain, I decided to build a ChatGPT project called ‘Prompt Library Builder’ that generates structured prompts on demand for 100 general purpose tasks in my voice. It took 15 minutes to set up. I’ve used it 60+ times since.

Here’s exactly what is in it:

This is where you define who the AI is in this workspace and how it behaves.

Mine looks something on the lines of this:

Why this works:

Compare that to instructions like ‘Help me write prompts in my voice’ and you’ll see the difference immediately.

I don’t paste context every time. The AI already has it:

File 1: signature-prompt-style.md
My 4-part framework for every prompt (Role, Task, Context, Output Format + Clarifying Questions ask)

File 2: prompt-card-template.md
The exact output format I want, keeping consistency of format intact

File 3: 100-Prompt-Systems-Library.md
Master catalog of all 100 prompts across 9 categories for general purpose work

These files ground the project in my system. The AI doesn’t guess what I want - it knows.

PS: I use Markdown files (.md) because it is the simplest code-block like format that LLMs can read clearly without losing context that could happen in rich text formats and makes it easy to store, edit, search and version.

After 60 uses, the project remembers:

I didn’t teach it as an instruction. It learned from corrections across chats.

Before projects, I was typing this every single time:

“I am a freelance consultant helping non-tech professionals learn AI. My audience is 35-50 years old, works in consulting/marketing/sales. Tone should be warm but direct, diagnostic style. No buzzwords. Give me 3 options…”

150+ words of setup. Every. Single. Time.

Here’s what I type now:

‘Generate a prompt for company deep dive for introductory client call’

What it outputs:

A complete and structured prompt with role, task, context, format, examples, and common mistakes. Ready to use. Zero edits needed.

That is because this is the workflow I want, this is the consistent output format I want and this fulfils my complete use-case with this project.

Before this project existed? I would have spent 20 minutes just getting one prompt after writing instructions, examples, formatting. Now it takes 2 minutes total for a high quality and complete prompt that is tuned in my voice.

That’s the ROI.

You set it up once. You use it forever.

Here’s the structure that works across any project:

Real example from my Newsletter & Linkedin content project:

Takes 10 minutes to write. Saves 15 minutes per session forever.

Don’t dump your entire Google Drive into a project. The AI won’t read it properly (or at all, depending on the platform).

Attach 2-5 high-value files:

For a Client Work project:

For a Content Creation project:

For a Coding project:

What NOT to attach:

Keep it lean. The AI performs better with focused context than drowning in files.

Mistake 1: Vague instructions
“Be helpful and professional” tells the AI nothing.
Fix: Specify tone in 3 words, give format rules, show an example.

Mistake 2: Uploading files but not using them
In ChatGPT, files don’t auto-load into chats unless you attach them.
Fix: Either attach files per chat OR use Claude (auto-loads project files).

Mistake 3: One mega-chat that runs forever
After 50 back-and-forths, the AI forgets early context.
Fix: Start fresh chats for new sub-tasks, stay in same project.

Mistake 4: No memory hygiene
Old mistakes or assumptions linger and poison new outputs.
Fix: Keep a “Decisions and Standards” file you update when something changes.

Mistake 5: Treating projects like folders
Projects aren’t storage. They’re workspaces with memory.
Fix: Set them up like you would brief a colleague - give role, rules, resources.

Not everything needs a project. Here’s my test:

✅ Create a project if:

❌ Stay in main chat if:

I have 6 most-active projects today:

Each project has custom instructions + 2-5 core files. Each saves me 60-80% of setup time or trying to give enough context.

I use both. Here’s when:

Use Claude Projects when:

Use ChatGPT Projects when:

The gotcha:
In ChatGPT, uploaded files don’t automatically load into chats - you have to attach them manually or enable “Reference Chat History” in settings.

In Claude, project files are always in context (unless you hit the 200K token limit, then it switches to retrieval mode).

I keep my Course Builder and Newsletter & Linkedin content creation projects in Claude (need quality + long context).

I keep my Prompt Library project in ChatGPT (need speed + iteration).

Use the right tool for the job.

Pick one workflow you repeat weekly.

Examples:

Then do this:

Step 1: Create the project (ChatGPT or Claude)

Step 2: Write custom instructions (use the template above, 10 minutes max)

Step 3: Attach 2-3 files (past examples, templates, or reference docs)

Step 4: Test it 3 times

Step 5: Use it every time going forward

By Week 2, you’ll have a system that produces 80% finished work on first try.

That’s the goal: not perfection, but consistently good enough to ship with minor edits.

Hope this inspires you to actively start using Projects to manage all your work / projects that need consistent context or is repeated or references the same files.

Do you want to build your Personal AI Operating System?

One of the things that we have been able to build over the last two years is a detailed system of workflows and agents that do a lot of our work with us.

Imagine never having to start a strategy or requirements documents from blank, you work LLM being completely grounded in your work context, not having to rely on your memory or intuition to be reminded of your to-dos and tasks, doing 90% of your work by speaking to your laptop, calls translating automatically to relevant tasks and having files-folders, Google search, media creation, learning all driven and managed by LLMs.

That is exactly what we have built for ourselves, and we are in the process of signing up the first 5 leaders for our 1:1 consulting where we will work directly with them to build this system. We have 2 slots open for our Feb’26 starts - so let us know if this is something you want to do with us by writing to us at: priyadeepsinha(at)gmail(dot)com

You can also reply to this newsletter email if you want to share something specific with us. Have a great week ahead and see you again next week!

See you next week again.

PD & Sonali
Work in Beta