How do you keep your prompts from getting just ramblings as you do the thinking?

It sounds like a fun way to "code", but I just can't imagine how it works in reality.

So this is the actual amazing thing. Claude does absolutely fantastic with rambling. As I am dictating I may change my mind three times as I keep imagining new cases or new constraints. And then at the end of my prompt I will say something like OK this has been a whole lot of brainstorming in different directions. I'd like you to summarize from me what you interpret as the final decisions and the reasoning behind it and the reasoning for why we are not going to adopt the proposed alternatives. And then tell me if you think I'm missing anything.

And it does. I don't even have to write out my final thoughts neatly summarized myself. Claude figures it out on its own and saves me a bunch of time. It's much faster for me to read it's logically laid out conclusions and confirm that that is actually what I had meant. Never once have I seen it get it wrong, but quite frequently it raises additional points for me to consider or highlights potential inconsistencies in what I have said.

So the actual session itself with Claude includes a whole lot of rambling and brainstorming, but then the actual documented artifacts of the session are the nice clean summary of what we figured out. And then, for example, when we move from a brainstorming design phase into a coding implementation phase, I'll start a brand new session so that it's context isn't polluted by the brainstorming and rambling but is rather starting off with the nicely organized document that had been created as the artifact from the previous session.

And this is why dictation works so well as a feature. Because it means I frequently wind up essentially "oversharing" my product requirements with Claude, and it will frequently suggest potential solutions that I had overlooked because I was already assuming that the solution would take a particular form. So it becomes surprisingly collaborative rather than me just dumping requirements into Claude.

Qwen 3.8 Flash Next also works very well with this, and frankly even the much smaller 3.6 35B-A3B did, too.

That's actually why this method is so powerful: ramblings are actually good as prompts. Seriously, try it.

I actually like this about talking vs. typing. often I feel like I get better results since I think give it more context around spec and decisions.

> I just can't imagine how it works in reality.

I can promise you it does, and it's amazing. I can stream-of-consciousness to the LLM, correcting and changing myself within the same paragraph of text and it works it out. Sometimes I clean up the text if I really went off into left-field but I just straight word-vomit into the text field and get amazing results. I don't waste my time editing, cleaning it up, etc, just speak and send. Dictation is gotten so good that I barely need to review/edit it. Even mistakes "committed" vs "commit it" and other mis-hearing are handling very well. Every once in a while the LLM needs to clarify but it's rare.

Less than 5 times out of thousands of prompts I've sent this way have failed due to bad dictation results and all but 1-2 of those were caught very quickly. In one, maybe two, cases it got pretty far along before I realized the mistake but I like those odds and nothing made it to PR (let alone a deployed environment) before it was caught. It was in reviewing the plan where I was like "what are you going on about?" and then scrolled up to see what it misheard from me which completely explained it's confusion and it going in the wrong direction. No problem, kill that session and restart with the right prompt.

As for "Fun", yeah, it's incredibly fun because I don't have to filter/edit upfront. It's how I talk to most LLMs, dump _as much_ context as I can verbally into it and let it sort it out. I, like pretty much everyone, can talk faster than I can type so I find myself giving MUCH more detailed prompts to agents as compared to if I had to type it all.

While keeping feature-parity, right...?

I think the memory hog you're thinking of is the official Bitwarden self-hosted backend. Vaultwarden is pretty light on resources.

Oh, one of the few blogs I was following. Sad to see this news.

Hope she keeps her pimp hand strong.

I am curious: Why is it bad to merge chat and work? In Claude I didn't have a porblem with it (although I switched to an OpenAI subscription shortly after they made the merge). Isn't the model capable of deciding if it needs the extra capabilities of Work?

The current benefit is that chat has no quota.

Which is probably why they will merge it. You can get a lot done in chat mode, especially if you connect custom MCPs to it. I can't see that continuing.

There is a quota, they just don’t tell you what it is and when it resets. You only get “you’re all out of pro messages, try again later”

Pro messages is like 50 per week for pro 100 and 200 for pro 200.

Regular chat, even on extra high thinking mode, is sort of unlimited. Idk at what point you hit the abuse gaurdrail but its really high whatever it is.

So is it 50 per week or unlimited, because it makes no sense unless chat messages are somehow not "messages"?

For a company with Open in the name, OpenAI is very opaque about things, there are all kinds of limits and quotas that randomly come and go, but they only show you the meters for a couple.

E.g. codex reviews come from codex quota but also has its separate quota they don’t tell you about. Same with security reviews which has another quota still.

So the chat messages can be sent at several effort levels and from various models including 5.6 and 6 pro. The 6 pro messages are limited and they don’t tell you how many you’ve sent or how many are left. But the 5.6 messages seem to be unlimited, however there could just be a much higher unstated quota.

Pro has a specific limit of weekly messages (50 for the $100 and 100 for $200 I believe)

The rest is basically unlimited unless you're doing something very weird.

Meanwhile I cant ask a simple question to Claude, not even to Haiku, because I used my claude code 5h limit.

You yourself also kind of pointed out why the discussion about USA and China is not off-topic. The niche Mistral wants to fill (afaik) is that it's European. I, as a European am really happy that they made such progress in so much worse (financial) conditions. I guess it's mostly geopolitics.

I'll be sure to check this out but in the meantime, I'd like to ask: Isn't it generally a good practice to run coding agents in a VM? It provides a security boundary so that the agent is restricted to the VM.

How does this compare? I see it's an "isolated sandbox", but what exactly does that mean?

Yeah when you self-host pipod, all of your pi sessions are running on your own VM. Within that VM, each pi session runs in its own container.

Not for long if the next release will be Gitea 290 instead of 29.

I have a feeling that nobody will die (most likely) for choosing forgejo or gitea... Apples and oranges...

I'm sure nobody will, but I disagree that it's an apples and oranges thing. But for you I'm willing to instead offer an example of a CNC machine making parts for... anything. The question and comparison is at its core still just as valid and relevant. Is the result way less important than "feeling good" about the solution?

But from the downvoting, announcing evaluation of these is contentious. Oh HN!

I suspect it's partly because people didn't jump from GPT-3 to GPT-6.1 Sol and partly because SOTA models from the last few(?) months have been able to tackle most of the regular tasks. It means this new model isn't different in that regard from Opus 4.8, if your mental benchmark is that they both are capable of implementing something like a CRUD app.