This feels like a useful read for somebody who's never used projects before in their AI tool of choice, but I think it comes up well short of helping you actually build a personal AI copilot. I say this mainly because projects have serious constraints in terms of the amount of knowledge you can actually upload to them, and the amount of knowledge required to be somebody's personal copilot seems like it has to be significant.
For example, I wanted to create a project in ChatGPT to house all of the images, articles, etc. that I came across showing interior design or architecture that I found beautiful. My hope was that by doing this for multiple years, I could develop a very clear perspective on what I value most when it comes to design, and use that to inform any future theoretical home building or renovation projects I would do. However, in actuality, my project could only hold 20 images, and that kind of killed the dream for me.
How am I supposed to create something that understands me intimately when I have such a hard limit in place?
First of all, that is a super cool use case and getting my gears turning.
You're absolutely right: the limitations around images specifically are very real and very frustrating. I have run into those limitations myself. I'll take it even one step further: There's plenty of moments where I wish I could just upload a video or audio.... then I remember that's not possible, and I first need to transcribe it.
Where the approach really shines today is in work/professional contexts where most of the context is text-based (or can be converted to text). In that specific use case, it's been game changing for thousands of professionals and PMs in particular.
It sounds like at least right now, for what you're trying to accomplish, projects are not the right solution. I really hope they soon will be because there's so much promise and potential in such an elegant feature.
That's just another way that AI's arrival is "uneven" (also described as the "jagged edge") though I'm encouraged at how early everything is, and how fast everything is moving.
If you're in a building mood and don't want to wait: LLMs aren't the only option out there for images. I've really enjoyed prototyping with Teachable Machine https://teachablemachine.withgoogle.com/ (this is mainly a classifier if I recall, and I'm sure there's way more out there that are better suited to what you're describing).
If an off the shelf product doesn't exist (but the capability does), your idea could be an opportunity to build for a use case that's not being served by the frontier LLM companies.
In the near term, if you have another context that is way more text-based where you can use this pattern that gives you value, it's a great way to start and build the habit and have a good intuition for when the capabilities increase. I've noticed they're increasing all the time.
Long term, I would encourage you to still save up those images because I don't see a future where what you're describing doesn't become possible.
Thanks for the reply! Hoping my use-case is possible in the future, too.
> Where the approach really shines today is in work/professional contexts where most of the context is text-based (or can be converted to text). In that specific use case, it's been game changing for thousands of professionals and PMs in particular.
I'm still not totally sure about this. The limit is based on the raw number of files, not the size of those files. In ChatGPT, that limit is 20, regardless of whether each is a tiny .txt or a folder of photos. In order for something to serve as an ongoing "personal ai copilot," it needs to evolve with you, which means continuously intaking more and more content.
Perhaps the workaround today is load up each .txt file with as much content as possible, but even at that point, the knowledge becomes stale the minute after you create it. How are you managing keeping this knowledge up to date given the limitations that exist today?
I'm looking forward to the day when I can actually have an ambient personal ai copilot that grows its knowledge about me as I just go about my day doing stuff.
Until that point, I'm just not sure that the current folder confines make it super viable for me.
I recommend giving it a hands-on try. I think you'll find the limits aren't as constraining as you think, especially as so much of the context also lives in threads that follow particular initiatives. I recommend using project knowledge for context that spans all initiatives, which grows at a slower rate.
If in practice you actually hit those limits for your project knowledge, your workaround is one very good solution (and unfortunate that it's needed).
I also recommend trying another LLM (Claude is the OG for projects, and they're increasing capacity over time).
And I love that vision btw - of all today's available options the closest to what you describe is definitely Cursor/Claude Code working on local markdown files. A few product managers and I will be doing some live demos of that on Monday and you're welcome to join us: https://maven.com/p/0a96cb/cursor-isn-t-just-for-coding-how-ai-native-p-ms-work
Here are several startup ideas inspired by the inefficiencies and opportunities highlighted in the article.
1. Dynamic Project Knowledge Sync Platform
A tool that automatically syncs and curates project knowledge from various sources (docs, Slack, project management tools) into your AI copilot’s context, eliminating the manual “human API” work of updating and maintaining context.
Why: The article points out the friction of manually updating project knowledge and the desire for a “living document” that stays current as your world changes.
2. Team Copilot Onboarding Layer
A SaaS platform that lets teams build, share, and continuously update a collective “copilot knowledge base” so new hires or team members instantly benefit from curated company-specific templates, frameworks, and lessons learned.
Why: The author dreams of new teammates starting with a copilot that’s “80% ready to rock,” leveraging team intelligence, not just individual memory.
3. Proactive AI Chief of Staff
An AI assistant that doesn’t just wait for prompts but proactively nudges users based on their calendar, project milestones, and recent activity—suggesting prep for meetings, follow-ups, or strategic focus areas.
Why: The article highlights the need for AI to “push” users, not just respond, helping overcome inertia and writer’s block.
4. Selective Context Management Tool
A platform that gives users granular control over which documents, threads, or initiatives their AI copilot considers in each conversation—enabling “selective context” for more relevant, less overwhelming AI support.
Why: Inspired by the wish for “selective context” like Cursor for coding, so users can mix and match relevant knowledge for each initiative.
5. Event-Driven AI Automation Marketplace
A marketplace of plug-and-play, event-driven AI automations tailored for knowledge workers (e.g., PMs, sales, support), focusing on “when X happens, do Y” triggers that integrate with existing SaaS tools.
Why: The article stresses the value of event-driven automations over batch tasks, and the challenge of identifying high-value, automatable workflows.
6. AI “Gossip” Recorder & Contextualizer
A lightweight mobile/web app that lets users quickly “gossip” (vent, update, or narrate) to their AI copilot via voice or text, which then distills and integrates these updates into ongoing project context.
Why: The “gossiping to your AI” habit is powerful but under-served by current tools; this would make it frictionless and valuable.
7. Copilot Knowledge Summarizer for Context Limits
A tool that automatically summarizes long chat threads and creates handoff documents when users hit LLM context window limits, preserving 90% of value in 10% of the space.
Why: The article notes the pain of context window limits and the need for seamless handoffs between threads.
A service that prompts users at the end of initiatives to reflect, then distills those reflections into “lessons learned” documents, automatically updating the copilot’s knowledge base.
Why: The author describes the value of growing project knowledge over time, especially with real-world lessons and retrospectives.
Long answer: Perplexity is a newer entrant on the "projects" front, and I haven't deeply tested it, but it definitely has all the elements (plus some Perplexity-specific stuff like being able to include links as project knowledge which could be cool for some uses cases)
I got 10 startup ideas reading through this article! Thank you!
This feels like a useful read for somebody who's never used projects before in their AI tool of choice, but I think it comes up well short of helping you actually build a personal AI copilot. I say this mainly because projects have serious constraints in terms of the amount of knowledge you can actually upload to them, and the amount of knowledge required to be somebody's personal copilot seems like it has to be significant.
For example, I wanted to create a project in ChatGPT to house all of the images, articles, etc. that I came across showing interior design or architecture that I found beautiful. My hope was that by doing this for multiple years, I could develop a very clear perspective on what I value most when it comes to design, and use that to inform any future theoretical home building or renovation projects I would do. However, in actuality, my project could only hold 20 images, and that kind of killed the dream for me.
How am I supposed to create something that understands me intimately when I have such a hard limit in place?
Hey Ned!
First of all, that is a super cool use case and getting my gears turning.
You're absolutely right: the limitations around images specifically are very real and very frustrating. I have run into those limitations myself. I'll take it even one step further: There's plenty of moments where I wish I could just upload a video or audio.... then I remember that's not possible, and I first need to transcribe it.
Where the approach really shines today is in work/professional contexts where most of the context is text-based (or can be converted to text). In that specific use case, it's been game changing for thousands of professionals and PMs in particular.
It sounds like at least right now, for what you're trying to accomplish, projects are not the right solution. I really hope they soon will be because there's so much promise and potential in such an elegant feature.
That's just another way that AI's arrival is "uneven" (also described as the "jagged edge") though I'm encouraged at how early everything is, and how fast everything is moving.
If you're in a building mood and don't want to wait: LLMs aren't the only option out there for images. I've really enjoyed prototyping with Teachable Machine https://teachablemachine.withgoogle.com/ (this is mainly a classifier if I recall, and I'm sure there's way more out there that are better suited to what you're describing).
If an off the shelf product doesn't exist (but the capability does), your idea could be an opportunity to build for a use case that's not being served by the frontier LLM companies.
In the near term, if you have another context that is way more text-based where you can use this pattern that gives you value, it's a great way to start and build the habit and have a good intuition for when the capabilities increase. I've noticed they're increasing all the time.
Long term, I would encourage you to still save up those images because I don't see a future where what you're describing doesn't become possible.
-Tal
Thanks for the reply! Hoping my use-case is possible in the future, too.
> Where the approach really shines today is in work/professional contexts where most of the context is text-based (or can be converted to text). In that specific use case, it's been game changing for thousands of professionals and PMs in particular.
I'm still not totally sure about this. The limit is based on the raw number of files, not the size of those files. In ChatGPT, that limit is 20, regardless of whether each is a tiny .txt or a folder of photos. In order for something to serve as an ongoing "personal ai copilot," it needs to evolve with you, which means continuously intaking more and more content.
Perhaps the workaround today is load up each .txt file with as much content as possible, but even at that point, the knowledge becomes stale the minute after you create it. How are you managing keeping this knowledge up to date given the limitations that exist today?
I'm looking forward to the day when I can actually have an ambient personal ai copilot that grows its knowledge about me as I just go about my day doing stuff.
Until that point, I'm just not sure that the current folder confines make it super viable for me.
I recommend giving it a hands-on try. I think you'll find the limits aren't as constraining as you think, especially as so much of the context also lives in threads that follow particular initiatives. I recommend using project knowledge for context that spans all initiatives, which grows at a slower rate.
If in practice you actually hit those limits for your project knowledge, your workaround is one very good solution (and unfortunate that it's needed).
I also recommend trying another LLM (Claude is the OG for projects, and they're increasing capacity over time).
And I love that vision btw - of all today's available options the closest to what you describe is definitely Cursor/Claude Code working on local markdown files. A few product managers and I will be doing some live demos of that on Monday and you're welcome to join us: https://maven.com/p/0a96cb/cursor-isn-t-just-for-coding-how-ai-native-p-ms-work
Here are several startup ideas inspired by the inefficiencies and opportunities highlighted in the article.
1. Dynamic Project Knowledge Sync Platform
A tool that automatically syncs and curates project knowledge from various sources (docs, Slack, project management tools) into your AI copilot’s context, eliminating the manual “human API” work of updating and maintaining context.
Why: The article points out the friction of manually updating project knowledge and the desire for a “living document” that stays current as your world changes.
2. Team Copilot Onboarding Layer
A SaaS platform that lets teams build, share, and continuously update a collective “copilot knowledge base” so new hires or team members instantly benefit from curated company-specific templates, frameworks, and lessons learned.
Why: The author dreams of new teammates starting with a copilot that’s “80% ready to rock,” leveraging team intelligence, not just individual memory.
3. Proactive AI Chief of Staff
An AI assistant that doesn’t just wait for prompts but proactively nudges users based on their calendar, project milestones, and recent activity—suggesting prep for meetings, follow-ups, or strategic focus areas.
Why: The article highlights the need for AI to “push” users, not just respond, helping overcome inertia and writer’s block.
4. Selective Context Management Tool
A platform that gives users granular control over which documents, threads, or initiatives their AI copilot considers in each conversation—enabling “selective context” for more relevant, less overwhelming AI support.
Why: Inspired by the wish for “selective context” like Cursor for coding, so users can mix and match relevant knowledge for each initiative.
5. Event-Driven AI Automation Marketplace
A marketplace of plug-and-play, event-driven AI automations tailored for knowledge workers (e.g., PMs, sales, support), focusing on “when X happens, do Y” triggers that integrate with existing SaaS tools.
Why: The article stresses the value of event-driven automations over batch tasks, and the challenge of identifying high-value, automatable workflows.
6. AI “Gossip” Recorder & Contextualizer
A lightweight mobile/web app that lets users quickly “gossip” (vent, update, or narrate) to their AI copilot via voice or text, which then distills and integrates these updates into ongoing project context.
Why: The “gossiping to your AI” habit is powerful but under-served by current tools; this would make it frictionless and valuable.
7. Copilot Knowledge Summarizer for Context Limits
A tool that automatically summarizes long chat threads and creates handoff documents when users hit LLM context window limits, preserving 90% of value in 10% of the space.
Why: The article notes the pain of context window limits and the need for seamless handoffs between threads.
8. AI-Driven Retrospective & Lessons Learned Generator
A service that prompts users at the end of initiatives to reflect, then distills those reflections into “lessons learned” documents, automatically updating the copilot’s knowledge base.
Why: The author describes the value of growing project knowledge over time, especially with real-world lessons and retrospectives.
Claude code with Cursor as a saddle is all you need
Tal, thanks for sharing this.
Can I use Perplexity's space feature to build my copilot?
Short answer: yes!
Long answer: Perplexity is a newer entrant on the "projects" front, and I haven't deeply tested it, but it definitely has all the elements (plus some Perplexity-specific stuff like being able to include links as project knowledge which could be cool for some uses cases)