iiinova — where your digital self keeps evolving

Joshua Crowley @joshuacrowley.com · 09/04/2025

Nice! How? SVG and some JS?

100

Joshua Crowley @joshuacrowley.com · 16/03/2025

“The question that most participants asked after the workshop was about the playground, and how they might continue this workflow.“ Almost a year later you still need to crack open a code editor to really try this stuff out.

000

Joshua Crowley @joshuacrowley.com · 16/03/2025

Here’s my reflections on a workshop I ran for designers on AI interfaces -> joshuacrowley.com/study/omnivo... I agree though, there’s not a tonne of thinking/writing about the opportunities for designers. But also we’re only just now getting models that are fast enough, cheap enough and good enough.

joshuacrowley.com

Joshua Crowley

100

Joshua Crowley @joshuacrowley.com · 13/03/2025

content-addressable filesystems! What a notion. I’m trying to reflect on how I use Cursor, as I consider files/folders, and how they work in the context of, well context!

000

Joshua Crowley @joshuacrowley.com · 09/03/2025

You should do it, simple way to start marketing whatever you’re building. Maybe a growing audience to shadow a project, if you’re vibe coding your own. Get insights and ideas.

Joshua Crowley @joshuacrowley.com · 06/03/2025

100% write more throw away code. I think we don’t appreciate how many more concerns even front end devs have besides just getting something to appear on the screen.

Joshua Crowley @joshuacrowley.com · 23/02/2025

Hi @jesseyuen.bsky.social added you to this handy starter pack, hope that's okay -> go.bsky.app/GXihHDK

Joshua Crowley @joshuacrowley.com · 12/01/2025

Looks good! Yeah you can get quite a clip going at the start, refactoring is a little slower. I found feature flagging really useful, because you can build out features so quickly, it’s easier to toggle combos of features then refactor with cursor.

120

Joshua Crowley @joshuacrowley.com · 04/01/2025

Yeah I guess I’m not aware of any library that handles end to end encryption for local/offline stored data AND remote data in any off the shelf package. Could be handy for the local first community.

Joshua Crowley @joshuacrowley.com · 04/01/2025

Yes I am! You can use the app locally without any auth. Or you can sync to a DO and I used Clerk to auth that. This repo by @codewithbeto.dev should show a pattern for that -> github.com/betomoedano/...

github.com

GitHub - betomoedano/groceries-shopping-list-app

Contribute to betomoedano/groceries-shopping-list-app development by creating an account on GitHub.

120

Joshua Crowley @joshuacrowley.com · 04/01/2025

Nice, do you have a link for the speaker component?

100

The reactive data store for local-first apps.

So in my app you can store your recipes, or your noughts and crosses games and all in a single data store thanks to @tinybase.bsky.social. User can keep it in localStorage or move it to the cloud.

tinybase.org

TinyBase

041

The UI component for each list is the source of truth for how the data works, it’s a contract between the user and the model. The workflow is prompt for component, which dictates the schema. User verify as meeting need. Generate system prompt based on the component to codify how to CRUD that data.

000

I’ve been experimenting with this for tinytalkingtodos.com you have a one size fits all schema, a row could be about a pantry item or a birthdate. Each row belongs to a list that contains a system prompt for an AI to help unpack it, and a custom UI to house it. It’s a lossy approach, a broadchurch.

tinytalkingtodos.com

Tiny Talking Todos

Less todos, more you.

220

There’s just so much detail to work through when implementing these experiences, it’s fun but also exhausting because they just pop up!

100

You could discard the audio tokens each turn, and inference off just the text transcription. But you’d probably want to be constantly re transcribing all the audio chunks to get the most accurate representation of the conversation, otherwise very lossy

100

How do you transcribe those tokens for the UX of chat? It’s likely a user needs a transcript to understand what’s been understood. Well you have to use another model/call to transcribe the audio input, like whisper. So you have a new problem, does the transcript/audio diverge?

100

Joshua Crowley @joshuacrowley.com · 30/12/2024

Multimodal models like Flash 2.0 and GPT4o can use audio tokens to natively reason over audio input. Neither can yet provide text and audio tokens that come from the same inference pass, and presumably wont ever. So that presents a flaw in these models no one seems to be discussing.

100

Would you consider doing the backend yourself? Using an IDE like cursor + Claude Sonnet 3.5, services like Clerk, Stripe, even Airtable can simply the requirements for the back end.

000

Joshua Crowley @joshuacrowley.com · 30/12/2024

You could build an AI party planner agent for example. Many steps like research, booking, RSVPs, catering, all executed on your behalf, with input. However for such a task, for many users the effort of planning is kind of the point, it’s an expression of an act of service for a loved one.

000

Joshua Crowley @joshuacrowley.com · 30/12/2024

I don’t know what agentic ai really means for users. I can see how an immediate barrier for deeper LLM experiences is access to user data. So will agents just be for integrating your service into a user’s auth/cloud/filesystem? a crutch for interoperability?

Joshua Crowley @joshuacrowley.com · 24/12/2024

TTT encourages you and your partner to sit down and work through a plan together. The Operator listens in, and makes notes adding todos to your lists as you discuss it in realtime. It cuts through the argy-bargy.

Joshua Crowley @joshuacrowley.com · 24/12/2024

Collaboratively plan all your last minute Christmas todos with your significant other and tinytalkingtodos.com. It's a local-first voice assistant with a focus on privacy and families.

Perfect way to coordinate your christmas todos

Gemini's Flash 2.0 is great, a true multimodal model I've been waiting for. Just finished upgrading TTT to use the new websocket live api today, just in time for Christmas. Here's a demo of tool callings and it's audio understanding. vimeo.com/1041500694

vimeo.com

Tiny Talking Todos

020

Although as I think out loud, Claude’s computer use fine tuning might help improve UX/UI iterations, given it can reason about UI elements. But still cost and speed prohibitive.

Claude is great at making working code, but has limited visual understanding in my experience. Needs extra prompting to get good UX. Gemini Flash/Thinking code gen probs ain’t 3.5 Sonnet quality, but might get there with visual understanding, speed and cost effectiveness. We will see!

Joshua Crowley @joshuacrowley.com · 21/12/2024

I’ll be trying Gemini 2 for code gen soon in my app, specifically to test its visual understanding. I’ve been pretty happy Claude Sonnet 3.5, especially because my current workflow is just to one shot via the users prompt. But excited to see where multiple fast iterations on a workflow gets us.

210

Yeah FWIW, biscuits.club inspired me to build my own local first app. Not easy to measure the long tail of open source.

Joshua Crowley @joshuacrowley.com · 20/12/2024

It’s still so early! I’ve spoken at dev talks where only a handful of folks have even heard of localfirst. There’s a big role to play in unpacking how you think about architecture with local-first -> tinybase.org/guides/the-b....

tinybase.org

Architectural Options | TinyBase

This guide discusses some of the ways in which you can use TinyBase, and how you can architect it into the bigger picture of how your app is built.

Joshua Crowley @joshuacrowley.com · 18/12/2024

One nice quality of building an assistant app is you can just ask for mock data to test your features, as you build them. "Can you try and delete a todo with a fake id".

000

Joshua Crowley @joshuacrowley.com · 15/12/2024

MCP has a page explaining how to best configure your LLM so it can pair with you on building out your implementation. First time I’ve seen this for an open source project.

modelcontextprotocol.io

Building MCP with LLMs - Model Context Protocol

Speed up your MCP development using LLMs such as Claude!

000

Sunil Pai @threepointone.bsky.social · 13/12/2024

Can’t wait to play with durable-scheduler looks so good

120

Reposted by Joshua Crowley

the todo app that... does are you getting it

4532

I’m following local-first principles, so probably. But it might be more ejectable, where you can take your data/lists with you into a mini version of TTT.

000

I also speak to user testing both, and how users wanted something less chatty! Which surprised me tbh. Here's a post with some background on TTT joshuacrowley.com/study/tinyta... if you don't have 40 mins :)

joshuacrowley.com

Joshua Crowley

000

Joshua Crowley - Tiny Talking ToDos

Here's a dev talk I gave on building voice agents in the browser at SydJS. In the talk I try and breakdown the different services you might use, and show how users can BYO keys, for a client based approach. I also demo using OpenAI's realtime websocket approach.

youtube.com

YouTube video by SydJS

100

Yep! The "Tiny" in TTT is a nod to TinyBase. Atm the UX is you select your collection, if it's Cloud > Durable Object. If it's "Device" just localStorage. So easy, just two different tiny stores.

140