Matt Dailey

this has been my experience too. once you work this way, it feels weird or sketchy to work through an entire change or issue or investigation in a chat session engineering is just (1) what decisions matter? and (2) make those decisions. then agents fill in the rest. our tools need to make key decisions clear to us, our team and our agents we're building in this space ref.tools as well. would love your take on our take! :)

I’ve been using RoughDraft almost exclusively since I got access. Undoubtedly an upgrade over TUI and linear chat for interacting with Claude Code. Musings: Linear chat as an interaction pattern is like trying to squeeze the universe through a tube of toothpaste. You’re losing so much when you reduce fable and Sol to linear chat. RLM patterns like dynamic workflows are uniquely suited to non-linear chat. Chat as a form factor kind of spilled out of the linear next-token prediction from a single context window, pre-agent era. But when you consider the global context as an object external to the context window and give a recursive agent access, you can abandon linear chat and other infinite-canvas-like form factors become more ergonomic.

@pgillian____ lol i have my own answer

@reactiverobot ...

software factory is the solution. what's the problem statement?

@grinich @bdougieYO @papercompute @evisdrenova @blackgirlbytes @EntireHQ @RhysSullivan @0xblacklight @humanlayer_dev @heymikasagi @2027dev @utpalnadiger I finally added a twitter handle for ref @ref_tools :)

Ok this is happening next week! 🪩 DEMO NIGHT 🪩 🗓️ Monday July 20th 🗓️ Featuring: • @bdougieYO - @papercompute • @evisdrenova @blackgirlbytes - @EntireHQ •@RhysSullivan - Executor • @0xblacklight - @humanlayer_dev • @heymikasagi - @2027dev • @utpalnadiger - @opencomputerHQ • @connorpato - @tembo • @pederzh - @manufact • @ShrekOverflow - game demo 👾 • @russell_h - @ConductorOneInc • @ChiefScientist - QueryGraph Stacks • @Alloutnikhil - NoInfra • @rjchint / @dudutwizer - Retriever • @akhileshrangani - buildy • @mikeclarke - @modemdev • @reactiverobot - ref•tools • @Chandrika633 - @quivlyai • @chelo_jimenez22 - @mcpjams RSVP here: luma.com/demo-night-s26

@grinich awesome!

@reactiverobot nice, would love to see the update. you're in!

We haven’t hosted a demo night at WorkOS in a while. Who wants to show the latest stuff they’ve been building? We’ll order pizza and drinks. 🍕

if the core atom of your work is managing a series of chats, you're always going to have the "move memory and context" problem. the solution is classic engineering: separate state from actions for coding tools this looks like a docs-first approach, you are context engineering in the doc/plan/spec with multiple agents and the agents become stateless but incredibly powerful actors. and as a bonus your human teammates can context engineer with you! that's how we think about it at @ref_tools. it definitely requires mental shift but after making it, relying on a single agents context window feels sketchy your last bullet, the idea that a single large plan is managed by multiple agents / models and you can track cost and usage is something we've been thinking about a lot too! but i'm curious, why desktop specifically?

would love a new desktop agent super app - lets me switch harnesses and models and multiplex across them - makes it easy to move memory and context - can orchestrate between models ( use Fable as a planner but a lower cost model for daily driver ) - can retroactively look at usage and optimize for cost / better results

i really dislike these for the amount of useless misinformation they represent on the web but they're necessary if you're not the household name I had to create one because people kept emailing me "how is different from ?" and i got tired of writing the same answer every time. if people are searching something about my product, i should have a page to answer it I ended up implementing evals for the competitor just to not have misinformation. but that felt super dumb. like why am I not spending this time making my product better? And in the end it got stale and they had to ping me to update anyways and i felt bad the extra sad thing is that these drive so much traffic that a tsunami of slop comparison pages is inevitable

The problem with these pages is that they almost always get things wrong, either intentionally or accidentally. It’s either dishonesty or negligence. There’s no real way to provide a realistic comparison because no product can win on all merits and no product maps with the same checkboxes or workflows. If your product doesn't have X, Y, Z but it has A that solves for X, Y, Z, and B, C, then you still lose because criteria was for X, Y, Z. Better to provide clear information and let customers make the comparison themselves.

You know a company is starting to lose the plot when they start publishing nonsense comparison pages. Who is juggling Linear, Jira, Rovo for AI, and Loom? 😂 (we still haven’t published a single comparison page)

the art with ai coding is balancing pre-work vs re-work

do you guys do spec/plan review? it seems like the main problem this solves for is human-human alignment on key decisions going into a change set. which is basically the entire point of reviewing the plan first. that said, "if you can't explain the code, i don't want to merge it" feels right lol

I just declared a moratorium against AI-written change descriptions (e.g. PR and commit messages, also issues/tickets) from my team. AI was writing change descriptions that were worse than useless to me as I tried to review PRs: outlining details of the code that could easily be seen by looking at the code, but omitting the higher-level framing needed to understand broadly what the code is doing. I think people like having AI write these things because the output looks structured and thorough, which makes it feel professional in a way. But this isn't actually valuable. Concise, high-level descriptions are better for everyone. If I need to use my own AI to interpret what your AI wrote then something is wrong. Let AI write code, sure, but for the description, I'd rather see your prompt than your output. We could maybe have extended agents.md with guidelines on writing descriptions, but this seemed a bit pointless since a good, concise change description only takes a few minutes to write -- not a significant time savings to delegate to AI. At least, it doesn't take long if you understand the code -- and if you don't understand the code, then I'm definitely not merging it.

yuuuup this is the way all engineering is going

we didn't set out to do this but what we've noticed from users is that @ref_tools makes the generalized "builder" role accessible in way that our prior generation of software eng tools do not

@Cloudflare approximately every single meetup I go to, someone pitch me on using x402 payments for @ref_tools cloudflare is probably the only entity with enough credibility to make me actually want to implement it!

We're opening the waitlist for our Monetization Gateway, which will allow you to charge for any web page, dataset, API, or MCP tool behind Cloudflare. The charges will settle in stablecoins over the x402 open protocol. cfl.re/4eUFdt6

@robinebers yup we're working on the local claude code plugin now i use cursor cloud but ref will track your local agents too

@reactiverobot that’s awesome! the problem i’ve seen for a few months now is that it’s incredibly hard to make people switch tools. even harder when anthropic says things like you can’t use your sub anywhere else. maybe think about how you can plug this into existing systems?

am still not sure about the cursor multitask feature feels like i need an inbox for tasks to keep track of what should have been done you using this?

@jnnnthnn 😥

nvm they’re nerfing it

Individual engineers are speeding ahead with AI while the org as whole isn't going as fast. I'll be talking about how to close that gap. This Thursday at @aiDotEngineer World's Fair!

this is very true for code tools right now! codex / claude code / cursor / ... are functions with the same signature: (code, prompt) => code and what's crazy is that AI is making coding look much more like design. engineers get to spend all of our time at the decision layer, thinking creatively about key decisions. and creative == collaborative The future of coding tools will be default multiplayer. It'll be tough for some engineers just like it was for designers when Figma became the default but in a couple years it will feel extremely obvious

the mobile product cycle drove everyone to become hyper-trained on network effects, app platform feedback loops, and Christensen's really good thinking on incumbents + insurgents now in the midst of the ai cycle and a widespread panic around moats everybody seems to have forgotten the power of networks and multiplayer systems you can’t reverse-engineer market fit but multiplayer ideas feel under-explored, because if you unlock one, a very very defensible system is waiting on the other side

as i go to more and more events, i start to get a sense for if a room is worth staying in the good rooms are the ones where i learn something and smile shoutout @dexhorthy for putting one of those together last night for @aiDotEngineer wf speaker prep and having had a sneak peek, i can say for sure you gotta see @vaibcode , @cbmenefee and @bdougieYO 's talks!

still long $FIG