Claude Cowork and chat are now one Claude
claude.com
claude.com
If you ask a thinking/research type question in 'Chat' versus 'Work' mode in these products -- say, something complex about politics, or do a multiturn business strategy, or want to work thru a new concept, you get very different answers.
The harness, steering, etc. in the chat/reasoning products is so much better for this type of question (that doesn't require code as a primary substrate).
Even as someone who mainlines like, 7 coding agents at all times, I regret that productivity fever will mean the regression of think-first-act-later AI UX.
All the sources of randomness are under the control of the provider, even if today's deployment structures mean providers introduce extra randomness due to the concurrent nature of the evaluation. Serialise the computation, feed it from a pRNG, and you have a fully deterministic result. But providers don't want to offer a deterministic result, and especially not one as fragile, expensive, and inefficient as a full serialisation would be.
†: For variants of "nothing" that include cost and deployment challenges.
Its also about very similar inputs -> different outputs. Even with everything you said, yes, same input would result consistently into same output, but sliightly different input and you might get completely different/semantic answer.
I already burn through Codex too fast.
The terminal sessions are running Luna on XHigh (on Pi) and GPT 6 Pro on ChatGPT.com is the main orchestrator.
Since ChatGPT and Codex have different limits, I can have 6 Pro orchestrate the Luna sessions. Using voice mode to talk back and forth about the software factory is really nice. Will be open-sourcing soon.
But Work gives the model a computer instance it can use for programmatic data access, API munging, etc.
Success cases I’ve had: chat - read the reports in my Gmail and put a counter proposal together.
Work - cross reference all the options with commute time using traffic aware map data. Ground it on drop off / pick up times from each option’s stated rules and hours. Create a table with commute and idle time break downs.
I must admit I spent a little time in ChatGPT this weekend, thinking I was in Codex. After some unexpected behavior (disappointed) I realized my mistake.
The model itself should figure out how much “power” it needs to run the query. I’m just asking a question. Whether you need to do a simple internet search or spin up a VM and run scripts is up to you.
I had this same thought recently. Is it possible to 'recreate' ChatGPT/Claude Chat easily in Codex or Claude Code? I assume it would roughly be just swapping the system prompt. Anthropic seems to publish theirs but I don't think OpenAI does
https://platform.claude.com/docs/en/release-notes/system-pro...
I tell chat to recommend me some books based on some of my favourites. Chat does it. I read some of them, liked some, disliked few. I give chat that information, it recommends me based on those, now I don't know which of the entries are relevant or not.
I don't want lock-in, so I try my best to tell it to generate copy-able MD syntax that I have to manually paste to an editor.
With Work mode, I tell the model to just maintain an append only markdown file, where it writes "still recommended", "not recommended any more", etc. beside titles. It does this after each turn.
This is just an example. Work - I see it as a mode that can maintain, update, and edit a file.
It's very convenient.
- This post
- The DeepMind Institute https://news.ycombinator.com/item?id=49727659
- Mistral X Mozilla: Private, Multilingual AI Browsing https://news.ycombinator.com/item?id=49723408
based on their account submission history they have a 25% hit rate (!) Seems very, very weird.
1. each one is a first-party vendor announcement, submitted very close to the announcement and under the vendors headline.
2. Titles are "verbatim" vendor headlines. No editorializing which follows HN guidelines to the letter.
3. Someone; a script/agent is watching vendor feeds and X accounts then submitting within minutes of the posting.
4. Altman/Navier–Stokes post got [duped] so prob not inside thing.
We will have reached HFK-ception when the bots are fighting to be the first to submit "How I used AI to build a High-Frequency Karma Farming Bot on Hacker News."
AI product updates probably tend to rank very well here. If you look at the domains the user is submitting, they’re pretty normal: twitter accounts of tech CEOs, big company blogs, etc.
If I saw a bunch of blogs with affiliate links or landing pages for slop products I’d be more suspicious.
Two years ago, I had no commute, and presentations were tedious ( i was NOT a good google slides user ), I kinda prefer this world for now
I see a version of that all the time and would never, ever do it.
one of the best parts of vacations is actually planning it! trying to find restaurants, organizing your days, trying to fit activities -- that makes you look forward to your time off.
really, i want to automate the boring parts of my life (did i really pay rent this month?) not the things that make me happy.
For some. Others hate or fear it. For others it just feels like work.
We all know the feeling of wanting to be successful, or to go on a great vacation. Showing these scenarios creates strong emotional response in the audience - we're now the people winning at work, going on a great vacation. This is especially important when you're selling something dull. See: insurance.
Claude: Would you like to go to Bermuda instead?
My github repos all have CONSTITUTION.md files that keep multi-agent dev grounded. Production runs trigger github actions that automatically check logs daily, file bug reports, etc. I launch Orca and literally just type "checking in" and the CONSTITUTION.md file governs the scanning of github issues and prioritization of which issues need to be addressed. Another process prioritizes and bundles them. When I type "checking in" the Orca worktree handles the launching of sub-agents that fix things. Occasionally I weigh in with an opinion or pick a recommendation. If I'm in the mood to pay attention I'll ask it to find another round of bugs and let's keep going. Often I just let it close after the first round. When I tell it "done?" it does a full regression and a production box review. Code auto-deploys to production twice/day. I don't regularly push to production manually.
On Fridays the CONSTITUTION.md kicks off a strategy and roadmap review when I type "checking in". As I have ideas I add them to the roadmap in one of the worktrees. Generally nothing happens until they're reviewed on Fridays together, unless I push one manually.. which happens.
Hermes runs on the production box. I have a few skills and quick commands that lets me check in on production runs and the status of things. It's read-only by design, basically my version of a dashboard. If I have a thought, idea, research link, or question it'll write to a github issue and we'll deal with it in the next check-in.
Orca lets me do all this remotely from an iPhone, and I talk to Hermes via Telegram. I generally don't INITIATE new work using either of these. I report issues/ideas and let the process do the things. But I can, and have.. I just tend not to. The whole thing churns, and so my inclination to "jump in and do a thing" is less than before. My inclination now is to toss things into the machine and let the machine work the schedule. I'm at a point where I could easily automate 80% of this and do my manual things 1x/2x week and I think progress would stay steady. I'll get there at some point, but I like the level of engagement I'm participating in now.
Occasionally I use Claude Code as a watchtower review of things, or do a wholesale code review, or do a review of logs. But deepseek is much better at building this machine -- I built something manually over time and through a ship of theseus process it got ugly. Deepseek reviewed the situation and wrote the CONSTITUTION.md and designed the processes independently. It favors deterministic scripts for process but launches them via LLM and monitors for exceptions; also, no memory system which also disqualifies Claude from being at the center of this. Memory plugins didn't work well because instructions start accumulating all through the chain of instruction files (AGENTS.md, memory, etc) and that gets very undeterministic very quickly. Instructions are written once, reviewed/audited intentionally.. it's best not to let LLMs learn and write their own dynamically (yet).
I've since used it as a template for other projects I've started -- they work the same. Currently building the GTM agent that will handle all SEO, marketing, keywording, etc for my projects -- it'll be a significantly autonomous Hermes agent. Don't really need Claude for any of this and my Deepseek bill is $50-100/mo.
"Claude, prepare me a presentation on XYZ."
I get to work, go straight to the meeting room, and pull up what it made to present.
It's barely coherent nonsense. Lots of irrelevant details, buzz words, wrong charts or confusing phrasing. Obviously LLM output.
I read it out.
When I'm done, I get a question about one of Claude's incorrectly inferred details.
The shame instantly kills me.
This is a scenario I've seen play out with coworkers. Except that last part, instead of dying or owning up to the mistake of trusting LLM output they waffle. Their shame circuit is broken.
"What does this bit mean?" "I dunno."
It was all awful.
I don't really know what the right response is, though. Walking out would be seen as way too hostile
Plenty of that even before LLMs too though.
This year, a good 40% of workshop slides are fully LLM generated, incomprehensible, and almost not matching what the speaker is talking about at all. Last year, while there were some poor presentations, standards were much higher.
Now that people are generating decks with AI, they’re basically worthless. I don’t need a bunch of bullet points you didn’t take anytime to think about on the screen while you talk about something you didn’t prepare for.
All the Apple Intelligence commercials were shamelessly this too. Felt like such poor branding for apple
If you are just presenting it for the first time after seeing it you are going to just be reading off the slides, which is a waste of everyones time.
I make restaurant reservations 7-8 times a year? My wife and I go on dates and you simply aren't getting a table where I live if you don't. Fyi we typically spend $120 or so on those dinners.
Most restaurants have a web UI which is very simple and easy to use.
At least in this city, reserving a table using the apps you already have on your phone is faster than using an LLM. Maybe these products are for people who know about the existence of Claude and ChatGPT but not Google Maps?
I suspect it’s because writers don’t normally have jobs and the only time they are ever in offices it’s to deliver a pitch for a project.
(They have iOS and Android apps too, but I've never bothered to install them.)
There's a lot in this launch, but the core idea is to simplify the product while giving users access to more capabilities. You no longer need to know ahead of time how much work a conversation might involve. If you're at your computer, Claude can use your local files and apps. If you close your laptop, Claude can keep working on its own computer.
This launch also lets you use Claude Design, Claude Docs, and Claude Slides directly from conversations. That's possible because we made Artifacts much more powerful: whenever Claude makes you an app, website, design system, or anything else, it can deploy an artifact with multiplayer features and databases.
As many of you probably know from your own work, giving users more power while making the experience simpler is really, really hard. It took many iterations to get to this version. We're far from done, but I expect people will be able to do much more while having to think about it less.
Just wait for the next leak!
Why/when would I ever need any of these? Can't y'all focus on useful features instead?
Not really proving their usefulness there :P
With AI safety top of mind so much lately, I can't help but notice the announcement does not address this.
With "Chat" mode, there was a user expectation that session had only limited capability to produce unexpected side effects, read sensitive files, etc.
With "Cowork" mode, it seems like more powerful capabilities have been on by default, requiring deep settings and safety understanding to disable if desired.
Merging modes feels like it is removing a simple and easy to understand risk management tool. How does the combined mode help users understand, manage and feel confident about what risks they are accepting?
I'm operating Team accounts, are there changes on how org-wide skill/plugin get loaded and used?
Specifically around the work UX. Trying to bring up scheduled tasks is really painful and required a lot of clicking around instead of just seeing the task results under my project. I hope this is being looked at.
Lastly, and this is a real nit. Let me turn off the "tips" when stuff is being worked on. I'm already paying, you don't need to force a rotating feature advertisement into the interface.
This was pretty frustrating. By trying to organise my chats, I actively made them worse. ChatGPT at least gives you the option to have either open or closed memory. Is this being considered?
Where do you see Claude doing better than what I'm seeing with Claude Code in these circumstances? Are there any reasons for me to leave my pretty heavily customized CC environment for the other versions?
Chat is a horrible interface for this, having to scroll up and down across lengthy conversations to try and pick up a thread.
I’m really sorry but I really fail to understand why this massive push to “chatify” everything. A task with its own context and multiple chats feel like a much better approach to work. Even if the work is spawned from a main chat window - eg “Claude we need to work on XYZ” and it creates a task to track this piece of work.
Further, recurring tasks again are really really hard to manage with a chat interface - which of the 10 chats has that question that the recurring task raised?!
And of course all this UX debt will likely stand in your way of building reactive items - ie spawning a task in response to some outside event. Think an inbound email being handled by a prompt that triggers a task and creates a draft response ready for my review and approval. Very hard to do with chat windows.
What Claude makes doesn’t need its own place either.
Enterprise admins choose when to turn them on
the frustrating part was deciding where a task belonged
What they’d started in one also didn’t carry into the other
You can choose how Claude checks in with you. By default,
Claude asks before taking an action. If you’d rather let it
keep working and check in only when something needs a closer look,
you can turn that on. You keep the final say."Sure! And since you were asking me about lobsters yesterday, I'll make it red and seafood themed!"
When casually a friend asks you to ask your Claude something about topic you chatted about earlier and then Claude brings back a secret you didn't want anyone to know yet.
Yes you can set up a project and then ask a question, but this is tedious.
There are several projects that injest coding sessions from lots of agents, but for the chat side I haven't found much, except replacements for Claude Desktop. Jan, for example.
They don't want a traditional UI with buttons and forms and labels because they want the interface to be "chat". The problem is that "chat" is tedious. And the turn-based, linear nature of the chat interaction model makes it even more tedious and unproductive.
The moment someone figures out a new modality for LLMs is when we’ll see the next hockey stick.
Honestly, that fills me with fear. LLMs exist to make money to their companies, and said companies are not gonna turn around and say, "you know what, go are going to make an android for each elderly person, that can not only help them with their medications, but that can actually make their medications, tailored to their biologies." Instead, they are going to go for the low-handing fruit of "you know Bob, the guy who makes jokes in meetings but who is grumpy about delivery timelines? Well, we are going to make an android to replace Bob. MetalBob will make even better jokes. The blue model will be able to explain in excruciating level of detail why timelines aren't reasonable. The red model will walk through the cubicles with a whip to ensure everybody keeps working all the time, and nobody goes to pee."
The downside of course is branching in got, but I usually don’t go of course more then a handful of tasks.
I had to engineer my own ticket management to keep opus on target. It’s been great for managing work, history, audit trails and commits are tagged with the task id.
I implemented subtasks to deal with the way Claude likes to stage its own objectives.
Then I made enforcement logic in the task manager so tasks can’t be closed out until reviewers have consensus on the same sha.
The adhd that is Anthropic demanded I build it and now I’m knocking out issues faster than Batman.
Often times I just sort of wish to undo.
>Each conversation is a Yjs CRDT tree. It can branch into sub-threads (recursively), and you can drill down, backtrack, edit, undo/redo, and inspect everything: tool calls, approvals, and the raw context JSON going to the model, etc.
It creates a new chat which has the conversation up to that point and a link to the conversation you branched it from.
The UI isn't tree-like, though. It's more like Git where each branch name is just mixed in with the rest of the soup of references.
Probably inevitable, but seems like a lot of disruption could happen there.
I know this all sounds abstract. I've been mulling over it for the past year and it's very hard; and LLMs are super janky and inconsistent so it's 100% not trivial. So in some sense I understand why a lazy bottom-of-the-barrel "chat interface" has become the de facto standard.
The protocol will of course be `text -> text` for the typical LLM (though some new models are structured, as we saw yesterday with Jev). Even so, the user doesn't need to be exposed to the protocol (almost ever). It's not like I'm crafting a POST request to send this form.
There's a lot of room for window dressing, and I look at AI like I look at the touch-screen. It was a fun technical toy until people (mostly Apple) poured in a lot of serious effort into "ok, how do we make this pleasant to use, intuitive, and genuinely useful?"
Dictation? Neuralink thought-to-text?
But I don't agree that they aren't product-ized. There are many applications doing calls to LLMs behind the scenes and are hits, leveraging structured data very heavily and not having conversations with users at all. I would say that there is a predictable scope creep from executives to surface a conversational aspect though. We need to bring representation to that so we can point to some other best practice to push back
API calls are expensive and slow, so I’ve shied away from integrating them into my personal projects. And of course there’s always the risk of hallucination.
Yes, they are expensive, but flexibility has a lot of value.
e.g. We went from somewhat smarter code autocomplete, to asking chatgpt copy paste, to cli agent running inside your project, managing session, to GUI to manage that, to projects where you talk to a "Chief of Staff" agent that manages other sessions, to who knows what's next.
I think the right interfaces for LLMs right now need to be very simple and easy to change/evolve. And chat still seems to be the best default solution.
Traditional and boring works for me.
I wouldn't call copy & paste code from a webui of chatgpt either traditional or boring. I'd call it tedious, error prone and guaranteed to get poor results. There is much better tooling and harnesses to leverage now.
Technology folks don’t really understand people and what they need.
This always happens. This is why woz needed Steve.
Steve Jobs is sorely missed tbh. For all the shit he got - he was a true visionary. He lived at the intersection of technology and the humanities… he kept preaching this. And now we are seeing why.
It sounds painful to debug
Honestly aside from the voice it uses you wouldn't notice a difference. Switching costs are low, vote with your wallet.
You need to use frontier models to understand where the puck is going, but also to use local ones for anything remotely sensitive.
Same, just a while longer ago.
I much prefer how OpenAI models write to Anthropic's, will probably revisit Anthropic in a generation or two. Context size is more limited, but no critical forgetfulness due to compaction so far, though I also like having plan files around both for future reference and improving chances of success at long form work.
Also tried out Kimi K3, was nice but slow (and apparently routed some requests to Claude anyways), GLM 5.3 was faster and still pretty good but the token allowances were kinda limited.
The harness is great, but opus is such an arrogant little prick that spews out unintelligible word salad. Opus 5 is so bad at communication it amazes me that somebody green-lit it. It's absolutely awful.
The fact that this isn't an acknowledged regression (and Fable 5.1 isn't much better) leads me to believe that people at Anthropic actually like Opus 5's output.
/balance-books hey claude here's the books. Go balance them
Claude then uses the guidance docs and various subagents to review the books, send off parallel workers with cheap models, then review with a more expensive model. Done in the repeatable, correct order with independent review every time. As new models come out, your workflow structure stays the same. You just delete some prescription from the guidance docs.
For example, I’ve spent the last few weeks building an internal ticketing setup for my agents so they stop trying to fix every tangential issue they encounter. I wrote a skill where subagents bubble out-of-scope issues up to a parent orchestrator, which decides whether to log a ticket.
Currently, that setup mirrors between a local SQLite database and an Obsidian vault: agents write to SQLite, which syncs to Obsidian; when I edit notes in Obsidian, it syncs back. Native Claude docs mean I could ditch the Obsidian middleman for remote review—editing tickets from anywhere while keeping the core SQLite state local to where the agents execute.
And unlike Sol: Opus like yoda, anything for unusable he is.
2) The first AI company to reduce their token pricing such that Astra/Fable can run for less than $150/day from credits would see a 5x usage increase and 20x spend increase from me. It’s currently around $500 for Astra and $5000 for Claude.
3) They seem to recently have addressed the fact that their safeguards were OTT so kudos to them for realising and now Fable is usable for wire and memory level programming.
The product management at Anthropic is pretty... nonexistent?
Claude's mobile app wouldn't show any Cowork initiated conversations, even if those conversations didn't leverage cowork specific features
I feel like Claude has some of the best UX, and hope this doesn’t dilute the experience.
Inside Claude code you’ve got cloud environments and local.
Let’s say as a normal desktop user you wanted to just automate something normal in your work, going to a website (maybe some internal app at your company) getting some data. Depending on what surface you used this will either not work, not work well (cowork), or work quite well but possibly be blocked because of bot controls (local Claude code) or again not work (cloud Claude code)
It was nuts. You can clearly tell these were different teams and products mashed into the same app. Super confusing for non technical users. Hopefully it’s a little better today
Genuinely curious.
I don’t talk to non-technical users that use the chatGPT app for work so I’m not sure if that’s been a positive change for them, but I would imagine if that was put in place for Claude it would be better for them. Chat is instant answers, different model, faster; work is more detailed, thinks more, has access to all the tools. This is what I see the users doing, but there is a third mode they use (deep research, exhaustively find me the right answer to a crucial question) that is underserved right now.
All of this is moot to me as a dev, I use terminal, herdr, pi… so I don’t see the jagged edges myself. But regular users do
Feels like the same "let us do _all_ of the thinking for you" messaging that Microsoft has with Copilot.
"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."