Agree on the fuzziness of natural language - but it would have to be implemented well - not as Siri today. The most notorious example, will be familiar to you, is the classic natural language issue with AND/OR. But it can be circumvented by additional questions. And it definitely should have failsafe mechanisms to prevent catastrophic consequences 
They say (and “they say a lot” to quote Pulp Fiction), that the new re-architected version of Siri will launch pre summer 2026, so we’ll see if Apple have finally got it right.
No voice assistant, be it Google Assistant, Alexa or Siri meet user expectations at the moment. But the likes of Gemini, Claude, ChatGPT, and Perplexity, are filling in for the moment with serviceable versions of a voice assistant. I find Gemini the best for my daily needs simply because Google know’s so much about me, and on that basis it provides solutions that are driven by personal context. I’ll add, that I also use Kagi for private search, the DuckDuckGo browser plugin which regularly blows away advertiser cookies, uBlock Lite, Privacy Badger and Ghostery browser extensions, so Google only know what I want them to know! 
I don’t use Gemini exclusively as a pseudo voice assistant. As I stated earlier in the thread, this is my all-in year of 20$ a month Ai products from the big players, strictly as a benchmarking exercise.
On iOS I tend to use Perplexity most often as it seems to integrate better with Apple’s core PIM API’s - calendar entries, contacts and so forth.
I top up my Anthropic API credits via OpenRouter right now, as Claude is night and day ahead of OpenUI and Gemini for programming workflows. But I don’t spend too much over my $20 a month entitlement as I get 50 credits with my Cursor Pro account, and you can spend your credits wisely with Cursor. The o3-mini (high) model from OpenAI isn’t as smart as Claude’s newer models, but it’s top of the middleweight tier on the benchmarks and is 1/4 the cost of the of a standard Cursor credit. And that makes o3-mini (high) the perfect partner to the newer Claude models.
I’d far rather a privacy focussed Siri as my prime voice assistant, as Apple also know a lot about my personal needs and interests, and are aware of my App preferences across their ecosystem to service those needs and interests (Siri will hopefully be able to access my Google data too via a permission setting). And from the looks of things, Apple are happy to hook up third party LLM’s (starting with ChatGPT) to Siri when you need to query a larger general knowledge corpus (that’s outside the privacy focused Apple ecosystem).
That’s a big promise, but I trust Apple more than most other Big Tech companies to understand that it’s worthless releasing something that doesn’t meet the standards we expect.
It actually can learn from programming language grammar - it just does a better job if it has examples. Just like an [intelligent] humor.
Claude 4 has a training cutoff date of January 2025 so it has no training on DT4. But you can prompt Claude by giving it the DT4 Scripting Dictionary and instructing it to write Devonthink 4 Applescript/JXA code using this dictionary - it does pretty well in that regard, including new DT4-specific features.
Unfortunately Claude is also frequently overloaded.
It´s actually such kind of hints, as to how to practically use things like Claude Code (and other tools) properly in DT contexts, or how to approach people knowledgeable in these matters (code assistant use as well as 'proper scripters) on this double background (presence of these tools; realities of DT, plus its ‘scripting culture’), that I would appreciate. And I think other parts of the community possibly as well.
Let´s not forget: the people driving this particular discussion here are 90% people with decades of coding experience, Apple/MacOS, webtechnologies etc.
I think it´s fair to say this is not really representative of the overall user base of DT. (Even if it were a majority, as one sometimes can get the impression, a community should still cater to these other parts of user base). It is in this broader context, that I use/propose terms like forum ‘culture’ and ‘protocol’.
The way the swath of users can adress all these challenges and affordances, here epitomized by Claude Code, should not be reliant on a myriad of one-to-one interactions, attempted ‘coaxings’, or experimentations, but there should be some attempt to synthesize this in some form of shared understanding(s), even better referenced in particular threads, guides etc.…
So, thanks for hints like this one…! 
1 Like
Do you mean via OpenRouter, Cursor, Claude Code, or Claude directly from Anthropic?
I can’t say I’ve bumped into that myself. But I tend to dabble with Claude late morning, early afternoon here in the UK. I have encountered problems once the US is open for business, all the way through to mid-morning the following day.
The Anthropic API. E.g. DEVONthink’s chat assistant performs an automatic retry in case of rate limiting but still I see an Overloaded error quite frequently this week (and especially today). Which is no surprise after reducing the costs for Claude Code that much.
1 Like
This is a specific area where Claude is far better than its competitors. Its ability to read the docs and make that part of its input tokens for context.
As in, making it a gateway drug to those on the $20 tier.
Even the Claude application is crapping out with overloads today.
But you can prompt Claude by giving it the DT4 Scripting Dictionary and instructing it to write Devonthink 4 Applescript/JXA code using this dictionary
Claude isn’t caching that as part of its dataset, so you’d have to repeatedly push the whole scripting dictionary each time you wanted to have such a discussion about scripting DEVONthink.
But DEVONthink supports Anthropic’s prompt caching which might reduce costs & response times. It’s of course best to separate static and variable stuff, e.g. add all the static parts (like the script suite) to a dedicated role and everything else to the prompt.
… similarly Elephas SuperBrains would come to assist here…
Other LLMs can use documents as a source of data i.e. on-the-fly RAG.
What’s different is that Claude is particularly good at coding overall compared with other models.
Claude is also very good at explaining code of most languages. If you are trying to debug/fix/revise code you can paste it and ask it to explain how it works or how to fix it or improve it in a given way - it is often spot-on.
1 Like
…and converts DEVONthink’s script suite definition (.sdef) very well to Markdown 
1 Like
If you have a paid Claude plan then you can set up a DT4 “Project” - that is a designated area which can retain specific documents to be used as part of your queries. So you could upload the DT4 scripting dictionary there.
Numerous other AI wrapper apps offer similar features - such as TypingMind “Agents.”
There is also an excellent app called DT4 which is in pubilc beta currently. You could upload the DT4 scripting dictionary there and then select / reference that dictionary when you open a Chat session and ask it to write a DT4 script. That works better than the built-in DT4 Scripting Assistant in my view.
That’s just a simple script.
And so by extension, you can add this to your CLAUDE.md file for project memory.
1 Like
Would you by any chance share that simple script for exporting DT’s script definition suite to Markdown, heck it can be vanilla plain text and I’ll use it to create my project Claude.md file.
Data > New from Template > AI > Script Assistant… is the script. For the conversion I just used a prompt and the first attempt already worked well enough:
Selected is the script suite definition (.sdef) of DEVONthink.
Convert it to a format suitable for prompts.
Ensure that all classes, properties, commands, elements and enumerations are included.
In addition, include the type of properties and returned values.
P.S:
I used Claude 4 Sonnet and Settings > AI > Chat > Usage was set to Best. But the latest internal build of DEVONthink 4 supports Claude’s extended and interleaved thinking, not sure if this made a difference.
1 Like
Thanks Christian.