Improving Chatbot Collaboration and Sharing
Key Points
- Nate announces his first deep dive into how chatbot experiences can be improved, presenting a personal “wish list” of fixes for the pain points he’s observed at scale.
- He stresses that open‑source LLMs now make it possible to prototype and launch new chatbot products in hours, encouraging builders to experiment, spin‑off, or even start companies.
- The core problem he highlights is that current chat tools excel at generating ideas but fall short at turning those ideas into actionable, ship‑ready work.
- The biggest usability gap is the lack of collaborative features—users can’t easily share a specific slice of a conversation, branch off into threaded discussions, or edit and comment together without cumbersome copy‑paste workflows.
Sections
- Improving Chatbot UX: My Wishlist - The speaker outlines the biggest pain points of current chatbots and shares a practical, actionable wishlist for making chatbot experiences less frustrating, emphasizing the rapid opportunities open‑source models create for developers and founders.
- Seeking Permalinks and Intent Shortcuts - The speaker laments that current LLM interfaces lack permanent links to individual high‑quality answers and require cumbersome, extensive prompting to capture user intent, urging the development of dynamic, citation‑friendly tools that streamline both referencing and intent clarification.
- Export Formatting and Task Visibility - The speaker outlines frustrations with current LLM outputs—broken formatting in exports, lost code blocks, and invisible action items—and calls for a middleware solution to produce clean, shareable documents and highlighted task summaries.
- Dynamic Pinning & Search Features - User requests automatic grouping, pinning, and contextual search tools to quickly locate, reuse, and highlight prior chat content—especially for HR hiring scorecards—without manual copy‑paste.
Full Transcript
# Improving Chatbot Collaboration and Sharing **Source:** [https://www.youtube.com/watch?v=z7LZlpr3PNQ](https://www.youtube.com/watch?v=z7LZlpr3PNQ) **Duration:** 00:15:28 ## Summary - Nate announces his first deep dive into how chatbot experiences can be improved, presenting a personal “wish list” of fixes for the pain points he’s observed at scale. - He stresses that open‑source LLMs now make it possible to prototype and launch new chatbot products in hours, encouraging builders to experiment, spin‑off, or even start companies. - The core problem he highlights is that current chat tools excel at generating ideas but fall short at turning those ideas into actionable, ship‑ready work. - The biggest usability gap is the lack of collaborative features—users can’t easily share a specific slice of a conversation, branch off into threaded discussions, or edit and comment together without cumbersome copy‑paste workflows. ## Sections - [00:00:00](https://www.youtube.com/watch?v=z7LZlpr3PNQ&t=0s) **Improving Chatbot UX: My Wishlist** - The speaker outlines the biggest pain points of current chatbots and shares a practical, actionable wishlist for making chatbot experiences less frustrating, emphasizing the rapid opportunities open‑source models create for developers and founders. - [00:03:54](https://www.youtube.com/watch?v=z7LZlpr3PNQ&t=234s) **Seeking Permalinks and Intent Shortcuts** - The speaker laments that current LLM interfaces lack permanent links to individual high‑quality answers and require cumbersome, extensive prompting to capture user intent, urging the development of dynamic, citation‑friendly tools that streamline both referencing and intent clarification. - [00:07:02](https://www.youtube.com/watch?v=z7LZlpr3PNQ&t=422s) **Export Formatting and Task Visibility** - The speaker outlines frustrations with current LLM outputs—broken formatting in exports, lost code blocks, and invisible action items—and calls for a middleware solution to produce clean, shareable documents and highlighted task summaries. - [00:11:25](https://www.youtube.com/watch?v=z7LZlpr3PNQ&t=685s) **Dynamic Pinning & Search Features** - User requests automatic grouping, pinning, and contextual search tools to quickly locate, reuse, and highlight prior chat content—especially for HR hiring scorecards—without manual copy‑paste. ## Full Transcript
I have spent a ton of time talking about
chat bots, chat GPT, Claude, Gemini.
I've also spent a lot of time giving
advice to builders who are trying to
work on embedding AI in their products.
But there's one thing I haven't done
yet, and that's what we're going to do
today. I have not talked about how I
would improve the chatbot experience. If
you are going to do a chatbot, how does
it suck less? because to be honest,
there's a lot of really big pain points
that are showing up now that we have
like a billion people on this app. So,
we are going to spend some time together
going through my wish list for how to
make this better. And why are we doing
this, right? Like typically what I do is
really actionable stuff that you can
take and apply right away. We are doing
this because you are not limited. We are
at a point with open- source models
where if you want to go and build this,
if you want to go and get a group
together and build this, if you want to
go hunt for a company building this, if
you want to found a company that does
this, it's open season. You can do that.
And it doesn't take that long. I saw the
first meaningful spin-offs from Chad
GPT's new open- source open weights
models in 6 hours. Now, not everybody's
going to be smart enough as a developer
to do it in 6 hours. I know I'm
certainly not. But the point is, if you
put out things into the universe, you
are more likely to get them back. And I
want to start talking honestly about
what I think sucks and what I would like
to see back, and I want you to talk back
to me. I want you to come back in the
comments and want you to say, "Yeah,
Nate, you were wrong here." Or, "Yeah,
Nate, that was great. Add that one." Or,
"Hey, did you think about this? This is
another one." So, I group these into a
couple of big categories. And the
overall theme is that chat is great for
ideas. Chat, GPT, Claude, others,
they're great for ideas, but they're bad
at turning those ideas into work you can
ship. Are they getting better at it? Are
they working hard at it? Around code,
yes. Around everything else, maybe not.
Let's talk about the pain points and the
fixes. So, the the first big categories
is sharing and working together. Right
now, it really sucks that I cannot share
just a relevant slice of the chat with a
friend and branch off of that and talk
together and come back into the chat.
There's no idea of threading. There's no
idea of collaboration. It's like the
chat is you and the AI locked in
together forever, right? That is not how
conversation and collaboration actually
works. Let me select messages. Let me
share a link to that part for view,
comment, or edit. I should be able to
start a new thread with that person
viewing, commenting, and editing that I
can tie back into the main thread. That
is not impossible to do. We have
invented editable tweets. It is possible
to do this right now. It sucks that
collaboration is just copy paste ping
pong. It's one of my biggest pains. I I
have joked with other leaders that I
feel like my entire day is arguing with
AI or arguing with people and copying
pasting back and forth between them. You
want to be in a place where you can live
co-edit and you can comment on answers,
see another person typing, get into
suggest mode. You want to get out of
this copy paste pingpong that puts you
into other things. I think we are seeing
an early hint of that in the way the
canvas is handled in chat GPT5. It does
not go nearly far enough. Neither does
clawed artifacts. Neither does the way
Gemini handles code. It's just not
enough. Part of why it's not enough is
that not everybody's a coder. And I
think right now, as much as I love my
development brothers and sisters, they
are getting the lion's share of the
attention. And people who do not code
need some attention, too. Right now, it
sucks that trying a new angle just
wrecks the original. There's no idea of
branching. If I want to start a new
artifact, the canvas just can't handle
that right now. The React component has
difficulty with that right now. It
certainly can't compare it side by side.
You can't merge stuff back easily. It
should not be that difficult to try a
new angle. That's how human innovation
happens. Right now, it sucks that I
cannot point someone to one exact
answer. I want to build a permal link to
one message that was a banger that
really got it right where like for
whatever reason chat GPT got it correct.
I want to be able to point to that one
answer and it's really easy and it's
really intuitive and then I can just
link to it and it's there for for like a
text message, right? And it's easy to
say what about this and I can give
attribution to whatever LLM I used.
That's fine. But I can't do that right
now. I have to share the whole darn
conversation with all the back and forth
that led to it and it's very hard to
find anything. Another big area besides
collaboration is how we get to intent.
Right now, I have to give you extensive
input on prompting, like pages and pages
and tens of pages and hundreds of pages
on prompting because intent is not
something LLM are good at understanding.
It sucks that you have to prompt write
for best results. I would prefer to type
the gist, get a clean request, and right
now, nobody really does this. Well, I
know there are prompt expanding tools.
I've tried them. They don't have
dynamically generated checkboxes that I
can confirm that walk me through the
variables. As a simple example, we are
at a point where LLMs can generate
dynamic interfaces like that. Comet
shows the way. Comet generates a dynamic
interface for sending an email. Great.
We don't have that today and we should
in the chat experience. I don't want to
have to prompt perfectly. I want to have
to prompt approximately and get a nice
clean request with little check boxes
that I can use. Right now, it sucks that
sometimes I want to be coached and
sometimes I want to take action and
sometimes I want the LLM to take action
and I can't convey any of that because
there's only one personality and the
personality seems to come pre-built with
the LLM and the personality adjustments
tend to be around temperament, not
productivity. And so I know that like
Chad GPT launched learn mode. I have
heard anecdotally that Chad GPT5 kind of
broke it cuz the personality is so
freaking eager. Please give me the
ability to adjust the degree of
agenticness for a particular project for
a particular request. If I want a
particular LLM to just go and work for
hours, let me turn the dial to 11 like
Spinal Tap. If I want to just have
little short loops and quick
conversations back and forth, let me
turn the dial down. It should not be
that hard. Right now, it sucks that
reshaping outputs takes forever. It is
so hard when I have something that's
about the right draft and I want to make
it more strategic or make it for the
seauite. How about you make it more
crisp? How about you use real examples?
Again, dynamically generate the
interface. Give me some one-tap options
and let me work through it. Don't make
me scream at the LLM and use my swear
words. I sometimes have to do that.
All right. Second. So, third big group.
So, we have sharing and working
together. We went through like five pain
points for that. We went through three
pain points for how you can convey
intent. I want to go through three pain
points for turning answers into real
work. It sucks right now that exports
break formatting. Right now, clean
exports to Docs, to Notion. They don't
keep code blocks intact all the time.
They don't keep headings clear. There's
all kinds of weirdness that happens with
copy paste exports. I will say Grock is
absolutely horrific about this. But
Grock is not the only one. Other LLMs
also struggle with clean exports across
this. We need something like MCP or a
middleware layer for exports. It just it
sucks so bad right now because you end
up spending more time fixing the
formatting than you spent getting the
LLM to say it. Right now, it sucks that
action items just disappear into the
chat scroll. Like, if it's if it's a
task due on Friday and you remind me,
don't just make it another chat. Like,
put it in yellow, highlight it, give me
a little calendar view of upcoming
tasks. Have some fun with it. The
chatbot is not the only thing right now.
It sucks that stakeholders stakeholders
don't want the messy thread, but they
have to get the messy thread if they
want to see how you work. Find a way to
give me a publish page. I want a
readonly summary of the best messages in
the conversation. It should not be that
hard. I'm really hoping someone from the
model makers is watching this. That
would be kind of nice. Another section.
So, we talked about turning answers into
real work. There were some sort of three
pieces there. I want to talk about the
idea of keeping it updated without
babysitting it. There's there's two big
pieces there, right? Number one, right
now it sucks that I have to remember to
rerun things. So, the scheduled tasks
feature in chat GPT is not enough here.
I want something that I can schedule
that will update dynamically as sources
change. Basically, if you're going to
promise me agents, give me smart agents.
If I'm going to let an agent do a thing,
give me more controllability over what
that looks like. Wow, the agent
discovered that the news source has
changed because the news itself is
shifting into a different vertical.
Well, then freaking update it and let me
know proactively. right now. It also
sucks that updates to code force me to
reread everything. It should not be that
hard to just refresh with differences.
Show me what changed since last time.
And that is true for chat and that is
true for artifacts and that is true for
code and a lot of other things we do. We
want to see the difference or the delta
and it just dumps everything. Another
section is trust and control. Right now
it sucks that I cannot easily see where
claims come from. there are these little
chips and I know that this is quote
unquote better than it was before and
sometimes it makes it up and doesn't
list chips. Give me something like I
have the receipts and it's a list with
sources and I should be able to copy the
markdown and frankly I should be able to
run a separate job on those sources to
validate that they're real URLs. This is
not that hard. This is literally what I
use Perplexity for all the time is to
check sources. Just build it in right
now. It sucks that the app remembers the
wrong stuff across chats. Yes, there's a
memory panel in chat GPT. There's not in
Claude. I want to see which memories are
being pulled into this chat and have the
option to edit them. I want to say, "No,
not that memory. Oh, add this other
memory. Make this a memory." Don't have
that right now. It sucks how sensitive
work feels risky. I know the New York
Times lawsuit and chat GPT is
particularly affected here because
they're being obligated to retain chats
they wouldn't otherwise. But even
leaving that aside, you want to be in a
position where you can have threads that
autodee. You can have threads that you
cannot share. You can have threads that
are locked to access in a particular
region. We want to have more options for
sensitive work and we just don't right
now. It sucks that I don't know the cost
or the time until it's too late. Just if
I'm in the API, continue to invest in
making those costs really clear. If I am
in the chat and I have a certain number
of messages on my plan, show me the
number of messages left. Make the usage
more clear and obvious. Another big
group is finding and retrieval and
reuse. Right now, it sucks how much
finished work gets buried. I would love
to be able to autogroup threads into a
project. Chat GPT, for goodness sake, is
able to know what related threads are.
Just auto group them together, pin the
best answers at the top, and highlight
them. Like I should be able to click a
dynamic little button saying this is
great, pin it or something like that and
it should just be really easy to find
finished work. Right now it sucks that I
have to reinvent paragraphs that that
that I've already written. I want to
pull a paragraph from an old chat that I
can link back and just dynamically pull
it in and say this is great more like
this. And it's just it's so much pain
right now. I have to dig through, copy
and paste it, put it. It's just it's
painful right now. It sucks that search
cannot find any kind of dynamic hiring
scorecard. So if you are using global
search, truly global search, you should
be able to dig into a lot of related
terms for your search. And so stay with
me here. You should be able to dig into
JD, job description, and scorecard all
at the same time. You should be able to
filter by person, by file, by time, and
come back. In other words, I am talking
about how we search within our chats,
not how we search on the broader
internet. If you have, in this instance,
you're in HR, you have a hiring
scorecard you built and you're digging
hiring scorecard thing. I want more
dynamic searches. I want it to know I
was talking about the hiring scorecard,
even if I didn't use those exact words
till late in the chat. Even if the name
is completely different and says
something like new strategy because I
happen to write those words in the first
prompt. I want global search that's
smarter within chat. We need that
because the number of chats we have.
Okay, thank you for sticking with me
through what is amounting to a lengthy
rant. The last section, quality of life.
Right now, it sucks that it asks 20
questions instead of just taking inputs.
It should generate forms for me more
often if it sees that I'm working
through something structured. This gets
back to intent clarity, but frankly, it
gets back to docs. What if I upload a
PDF and it's a government form? Can't it
dynamically generate the form and let me
like fill in little pieces and give me
advice as I fill it in and then I can
finish it up in like one click much much
easier than asking me 20 questions in
line. Right now it sucks how much tone
whiplash wastes edits. The tone presets
are really clunky and they overindex way
too fast and it's so painful to edit
back and forth. Give me better dials.
Give me more nuance in my tone presets.
which may come back to the idea that we
just need a better LLM for writing and
we don't have great options right now.
And I know that people are raving about
Kimmy K2 and writing and all of the
vocabulary that it uses. It's not about
vocabulary. It's about these kinds of
breadandbut features that enable you to
actually make the writing work for you.
And speaking of whiplash edits, you
should be able to find a fingerprint
tone that you like in your LLM and
preserve it and save it. It should not
be that hard. Right now, the best
options we have are these system
instructions, which feels like beating a
screw with a hammer. Like, you can kind
of make progress, but it's still more
painful than it should be. Last thing,
right now, it sucks that I cannot roll
back to the good one from this morning.
Version history and quick restore on
chats, like we just don't have it. If I
go into the canvas, I can't roll back to
the good one. If I go into claw, the
best thing I can do is click on a
previous version. There's no sense of
roll back or checkpoint. That's really
frustrating. If you want to look at the
overall theme of what I've talked about,
what I'm asking for. I want to make chat
sharable. I want to make it shippable. I
want to make it more self-maintaining. I
want to be able to share slices and
expand intent more easily. I want to
reshape and export. I want to auto
refresh diffs, show receipts, control
memory. Really, what I want is to jump
from a clever chat to an actual
workbench. And I don't see nearly enough
motion on this 3 years in to the LLM
revolution. Like we're still in the
chatbot and people are saying builders
need to build apps that aren't just chat
bots. Well, we do and I talk about that
a ton. I'm all in on that. But we also
just need to use better chat bots and it
shouldn't suck this much. Thank you for
staying for my rant. Let me know what I
missed. Let me know what you disagree
with. What would you like to see changed
about Chuck?