EDIT: Ah dammit. Excitement got a head of me...word replacement, not work!

I just found a genuinely awesome use case for an LLM. At least for myself, though I'm sure I'm not the first person to discover it. It's incredibly simple, and I feel like this community would appreciate it more than any other.

I use STT apps constantly and love them. VoiceInk, Spokenly, and Superwhisper locally. Parakeet is incredible. I also use built-in AI and LLM post-processing on them afterward for correction and cleanup, but one of the things that gets annoying is correcting certain words that the local model always transcribes incorrectly.

All of these apps have "Word Replacements" and user dictionaries. You can set it so if the transcribed word equals x, it rewrites it as y. That happens programmatically, not through the LLM, so it's guaranteed to work.

Spokenly lets you import a list of these replacements and export them as JSON for backup. So I had an idea.

1. I created a few replacement word entries.

2. I exported them to JSON so I could see the schema.

3. I fed that JSON to GPT 5.5 and told it to learn and understand the schema, then come up with a plan for creating more entries in a way that would successfully import.

4. It figured it out on the first try, and Spokenly successfully imported the single word replacement JSON. Now Codex knows how to use this system.

5. I gave Codex this prompt:

> Go ahead and read through all of our chat histories, perform a word frequency analysis, look for words that are uncommon, such as proper nouns, abbreviations, and the names of technologies, and create a list for me to pronounce into my STT tool, one after the other. Then, take the transcript in the message I send you back, assess it, and turn it into a word replacement JSON, using the word I pronounced and the mistranscribed word, which will be in the same order as your list.

It scoured through all of my personal context, since I keep my own memory system in the form of Markdown files. One by one, it started pulling up all of these various names, technologies, and other things I'd mentioned, along with a bunch of extra material. It gave me a numbered list, so all I had to do was toggle my speech-to-text, sit there, and go straight down the line pronouncing each word once in the order provided. Whatever it transcribed, Codex knew was what the STT engine thought I had said because they were in the same order. Anything it was unclear about, it asked me to clarify.

I imported it, and holy shit...this worked so freaking well. It generated this gigantic list of word replacements and even added certain variations. For instance, I talk about the macOS MPV-based player called IINA quite a bit, since I have some skills and scripts related to it. Every STT tool always transcribes that as "ena," "Ina," or even "Enuh." Codex covered all three of those spellings for me and made them three individual word replacements. Now I can safely say "IINA" into STT without ever having to worry about correcting it.

Example: https://i.postimg.cc/v86QMsMD/Clean-Shot-2026-07-12-at-21-46-33.png

I'm so damn happy this worked because I was already a huge user of speech-to-text, but now I use it even more.

Then I went one step further and created a skill called stt-fix. It lets me use the syntax open code -> OpenCode at any point in a chat and say nothing else, and it instantly generates a JSON file with the word replacement for me to import. I can do as many as I want at a time. I've even had it review entire conversations and tell me which words it thinks might have been mispronounced but usually ignores because the built-in LLM post-processing knows what I meant to say. For instance, without replacements, the local model often hears "LOMs" when I say "LLMs," and Codex found that itself. I just review the list, and boom, I'm done.

Then I found out that the Superwhisper CLI can add these replacements directly from its command line, so I switched to that. Now, at the end of each conversation, I just have to trigger and approve and it just does it in th ebackground for me, no clicking required (Spokenly's CLI doesn't do this...yet?)

I love use cases like these. I already loved using speech-to-text apps, and now I love them even more.

Try this workflow out. You'll love it!

Hey there, I wanted to provide some feedback on the new Raycast V2 Beta which I just got into today.

Version: 0.58.1.0

MacOS: 26.4.1

HW: Macbook Pro 16, M5 Max 40c, 128GB/4TB

Monitor: Alienware AW3821DW

Raycast Subscription: Pro (w/o Advanced AI).

NOTE: I know there's currently some impatience and frustration from some Raycast users here, and I don't want this to turn into a thread where everyone just piles on hate. I'm leaving this detailed feedback because I love this app and I use it dozens and dozens of times a day. So please be constructive to the Raycast team if you are going to reply to this thread! <3

Visuals

• Font rendering seems to be off - comparing the two, the original's font seems to be a little thicker and smoother
  • In the OpenAI Codex app, when they recently added "Native macOS Font Smoothing", I found that I actually preferred it off. Maybe that's the same thing here?
  • In particular, notice how much worse the "Open Command" text looks compared to the previous one.
    • Old: https://i.imgur.com/f7ECU4Q.png
    • New: https://i.imgur.com/SjQrVeW.png
• Please let me disable the bottom fade-out - I'm sure it's aesthetically pleasing and in line with Liquid Glass, but given the limited height of the Raycast window, I don't want to obscure any of it
  • If it's about making sure the user is aware that there is more to scroll, I get it, but also, this is Raycast...I will assume there is more unless I start typing and it shows me otherwise :)

Onboarding

• Migration to the new one felt pretty seamless. For those who don't know, it walks you through everything in one step and disables all previous shortcuts:
  • https://i.imgur.com/1MXOw3I.png
  • https://i.imgur.com/54uSc1V.png
• Had 70 extensions, it would have been nice if the setup let me continue through while extensions continued downloading instead of having to wait. For a one-time migration, really not a big deal, just throwing that in there!

Performance

• Double-tap CMD to open window switcher is unreliable - quite a few instances where it doesn't launch, or it launches delayed
  • This is a massive problem for me - my muscle memory has taught me that I can rapidly type "CMD-CMD + iter" and press enter to instantly get to iTerm
  • Lately, I've been accidentally typing "it" into dialogs and text editors
  • When Raycast does trigger, because of how long it takes, it misses the i and starts typing ter - because of the fuzzy searching this means that sometimes Telegram gets triggered instead of iTerm
  • I have changed over to Switch for my window switcher needs now.

Extensions UX

• I understand that being able to fully manage and change extensions settings straight from the launcher including shortcuts is a newish feature (which is nice!), but navigating settings for Extensions feels like it's gotten heavier and more clunky. Sometimes I want to do some extended amount of tweaking to my extensions, learn what the commands that come with an extension do, etc.
• The settings window needs to be resizable, I think it's too constricted for this newer less dense layout
• The lack of searching through my extensions all at once with the original filter bar is a downgrade (both built-in + custom extensions)
  • I understand why it made more sense to separate out Shortcuts into its own dedicated menu, but if I want to quickly find an extension like Antinote, I have to go scrolling for it
• I think I preferred the old method of extension management, where extension settings, aliases, and hotkeys were all in the same "table" which operated under the same "filter"
• I have an AI Command called "Fix Spelling and Grammar". Yes, I can go to Shortcuts → CMD+F → type "Fix" to find it, but the only thing I can do from this menu is change alias or shortcut: https://i.imgur.com/N3LlaW5.png
• In V1, I could hit a button and directly edit it there: https://i.imgur.com/R0vdRTA.png
• In V2 I need to go to AI, then Commands, then find my command manually (since there's no search here), and once I do, I have to tap the 3 dots (which is a tiny, tiny little button) to get to the "Edit" part: https://i.imgur.com/UDqjx0B.png
• If it was up to me, I think I'd create a unified "Control Panel" tab instead that has fuzzy search, where fuzzy search searches through every single built-in Raycast option title, both built in and third party extension title, description, command list, and command description (important one), and allow for managing shortcuts, aliases, and invidiual command settings all in one go from here.

Dictation

Really not having a good experience with dictation. When I go to set the shortcut, I press left OPT once. It shows up as Option, and then continues to cycle through its little tutorial that shows me I can single tap, double tap, or use a sequence. Only a double tap would register?

https://i.imgur.com/a8HtFt3.mp4

Feedback & Community

The built in App Feedback tool is a bit frustrating.

1. Category/Sub-Category picker is extra work to do each and every time. I'm sure this is routing to some ticketing system on your backend, so I understand why it's there.
2. Some categories don't exist. I didn't see a "Dictation" category (presumably because Dictation is a beta feature and you're reusing the stable feedback tool).
3. The window is not resizable.
4. The file picker doesn't allow for CMD V to paste in screenshots, which is another stepping stone in the way of feedback

I just want to throw this out there: One of the things that AI is phenomenal at is sorting and differentiating between data...I'd consider utilizing a cheaper LLM model that is fully index'd/RAG'd against the Raycast documentation, and then have that LLM classify and sort the feedback automatically based on the context of the feedback. No idea what your backend looks like or if this is even reasonable, but I'm just throwing it out there.

The reason I'm posting this here instead of with the feedback tool or Slack is for a few reasons:

1. I didn't want to repeat that over and over and over again for multiple pieces of feedback.
2. I didn't want to create a Slack account...I understand that's where you started, but Discord seems like a much better option. Many apps and developers are on Discord already, and I think you'd get a lot more feedback if you opened up discussions there. I'm sure you'd have Raycast ambassadors who would love to volunteer and help moderate. But I understand this means more work for the Raycast team, so I'm not really blaming you.

These are growing pains of a V2 rewrite. I want to make it super clear again that I am not looking to just pile criticism on the team. I just have found this to be such an essential app that really care about where it progresses!