@lu__jasper i like to think my twitter addiction was an inspiration
English
Vincent van der Meulen
4.4K posts
Vincent van der Meulen
@vinvan
co-founder @mainframe. engineer, semi-retired designer. prev: @figma @diagram (acq) @facebook.
NYC Katılım Ağustos 2013
1.8K Takip Edilen9.4K Takipçiler
@imjaredz 🙌 same!
feel like middle managers are going to take off soon — it’s one of the only ways to stay sane
Vincent van der Meulen@vinvan
fable produced 60+ (!) ready to merge prs for @mainframe overnight. here's how you can set up a software factory using fable as an orchestrator. prompt included! 1. narrate your *entire* todo list — everything that's top of mind — using dictation. i used @usemonologue and my prompt was ~10K words! you have to use voice because we want excruciating levels of detail. 2. locally, tell fable to plan all the work and put it in @linear. the goal is to make it such that a dumb model dropped in linear knows *exactly* what to do. aka all issues should have a clear DAG, plan, and optimize for max. parallelizability. i recommend running fable on high and having it fan out to codex subagents on xhigh (using the cc codex plugin.) to save usage. just make sure that fable itself is doing the architecture for the hardest issues! 3. now switch to @DevinAI. we'll use devin to run our factory because our orchestrator needs to be a cloud agent (it's going to run for a long time!), and be able to spin up other cloud subagents. 4. hook up the @linear @SlackHQ mcps. this will let your agents read the fable planned work, and communicate with you/your teammates. e.g. it can talk to your coworkers to make sure it's not stepping on anyone's feet! 5. a good feedback loop is *crucial.* before starting, make sure you have a bug bot (we use @cursor_ai.) and that your agents can test all its changes. e.g. we use the fantastic @limrun to make sure our agents can do ios development e2e. 6. create a devin ultra (fable) agent — this will be your orchestrator or 'middle manager.' we're going to tell it to look at linear and dispatch subagents for all tasks. the middle manager's goal is to keep spinning up subagents with the right level of intelligence until all the work is done. and keep all agents unblocked while you're sleeping, hanging out in the park, whatever! the reason the middle manager is able to orchestrate all of this is because fable planned everything in linear. and because the middle manager itself is also fable (contrary to most of its subagents.) 7. to prompt your middle manager, attach the middle manager prompt i'll post in the comments, modified for your repo. it's important that we *attach* the middle manager's prompt to our conversation vs. typing it out, so it never gets compacted. the middle manager's context window is holy. so our prompt tells it to spin up subagents — instead of doing things itself — for pretty much everything. 8. the prompt has some other cool, non-obvious details. firstly, every agent is told to keep working until it's met the burden of proof. in our case that's a (self)-review, bug bot being green, bug bot fixes being elegant, and a recording. we tell our middle manager to enforce this. secondly, we ask the middle manager to make sure linear stays up to date and communicate via the slack mcp. specifically, it should tag coworkers in slack whenever it has a question and post heartbeat updates in the eng channel. 9. your software factory is ready to go! in our case, it worked for 12+ hours without getting blocked. and produced extremely high quality prs. addendum: all of this is possible because with @AnthropicAI fable, models are finally smart enough to orchestrate/navigate large sets of work and their dependencies. and fable is a great software architect! now of course the 'downside' is that you're now review bottlenecked. we have 60+ PRs to review! but i think as an industry we should be able to come up with some innovation there. we'll certainly try our best to help here with @mainframe too! hope this is helpful! happy to answer any qs.
English
"The Boss Agent"
With Fable & Sol I think we have unlocked a new and better way to run agents. A single boss agent.
Fable is really slow so I find myself often with 10+ parallel agent threads all working on different tasks.
The problem is that once I merge one PR, the rest get out of sync and conflicts need to be resolved
The solution: A Single Boss Agent
I now do all my work through one single agent thread. It handles spinning up child agents, coalescing results into PRs, and fixing conflicts. It can message out to all threads or babysit them for you.
Fable is an order of magnitude better at delegating than Opus.
I can see this becoming the next UX paradigm for agents. Just one simple thread.
English
code review is *unnecessarily* difficult today because github prs don't require chat transcripts/model metadata.
it's impossible to answer:
- "which model wrote this pr?"
- "is the pr done or still a draft?"
- "has the model done due diligence?"
would love to see improvements!
English
ok! quick updates from me:
- i left anthropic (who does that?!)
- had a baby (i love her)
- and am joining @cursor_ai as head of design (eep!)
it's been a low-key dream of mine to nurture a team that cares so deeply about craft, quality, and building great tools. very excited!
English
@levelsio idk if you’re looking for more products but i’d pay for this!
English
So I am trying to hit a 500 calorie deficit every day + hit my protein goal of about 150g (2g per kg bodyweight), and I used Claude chat for that, but after a week it starts losing track so I asked it to export its data as CSV and copy pasted that into Claude Code on the VPS
And asked it to build a little calorie tracker called 🥩Caltrack with a Telegram bot too, so I can log whatever I eat or drink in there either via Telegram or if I want more granular via Termius SSH on the VPS
I like to do this mix of Claude Code and a dashboard/chatbot because you can ask more deep questions to it like "how am I doing", "what to improve" etc. it's just much smarter and "able" on the server than the Claude chat app by itself
Anyway my goal is to get @marclou levels of body fat, which is LOW, right now we're strong and lean but we can be leaner, let's try 😊😊😊
80% of body fat is decided by what you eat, only 20% what you exercise, some say even 90% vs 10%, I agree, I eat very clean but I was mostly staying on maintenance
Silly calorie dense food like yogurt with low protein, taste nice but it gets you away from your calorie deficit fast, and to figure that out you kinda really gotta track things
Other people suggest crash diets like eat only sardines, we know from studies though that crash diets are temporary, you just gain it back
Claude says the ideal is just 500 calories deficit per day and hit your 2g/kg protein goal and you maintain muscle and reduce body fat! So I'm doing that
And this app I made helps a lot!!!
Thank you for your attention to this matter!!!
English
@geoffreylitt haha nice i do this too!
also works for runs, which makes it a nice way to get some exercise in
English
A workflow I'm enjoying: "Walk-driven development"
> go on a nice walk outside 🚶
> record a long audio note: ideas, goals, things to build 🎙️
> agent auto-creates docs/tasks, and kicks off cloud coding agents for me 🤖
English
i tried giving a fable agent various writing assignments, and having it optimize its submissions against the @pangram API, to produce low-slop writing.
On a simple prompt it got 0% after 11 rounds of iteration, and on a hard one it failed - and did a crazy reward hack instead:
English
@Infoxicador not always what you need, but good for when you e.g. have a two week sprint to rework a product in a certain way!
aka you roughly know what needs to happen, but are going to lose a ton of time doing it sequentially/tackling ~3 tasks at once
English
Is it just me or this looks good in theory but over complicated and in effective in practice?
Maybe useful for 0 to 1 and prototype products?
Vincent van der Meulen@vinvan
fable produced 60+ (!) ready to merge prs for @mainframe overnight. here's how you can set up a software factory using fable as an orchestrator. prompt included! 1. narrate your *entire* todo list — everything that's top of mind — using dictation. i used @usemonologue and my prompt was ~10K words! you have to use voice because we want excruciating levels of detail. 2. locally, tell fable to plan all the work and put it in @linear. the goal is to make it such that a dumb model dropped in linear knows *exactly* what to do. aka all issues should have a clear DAG, plan, and optimize for max. parallelizability. i recommend running fable on high and having it fan out to codex subagents on xhigh (using the cc codex plugin.) to save usage. just make sure that fable itself is doing the architecture for the hardest issues! 3. now switch to @DevinAI. we'll use devin to run our factory because our orchestrator needs to be a cloud agent (it's going to run for a long time!), and be able to spin up other cloud subagents. 4. hook up the @linear @SlackHQ mcps. this will let your agents read the fable planned work, and communicate with you/your teammates. e.g. it can talk to your coworkers to make sure it's not stepping on anyone's feet! 5. a good feedback loop is *crucial.* before starting, make sure you have a bug bot (we use @cursor_ai.) and that your agents can test all its changes. e.g. we use the fantastic @limrun to make sure our agents can do ios development e2e. 6. create a devin ultra (fable) agent — this will be your orchestrator or 'middle manager.' we're going to tell it to look at linear and dispatch subagents for all tasks. the middle manager's goal is to keep spinning up subagents with the right level of intelligence until all the work is done. and keep all agents unblocked while you're sleeping, hanging out in the park, whatever! the reason the middle manager is able to orchestrate all of this is because fable planned everything in linear. and because the middle manager itself is also fable (contrary to most of its subagents.) 7. to prompt your middle manager, attach the middle manager prompt i'll post in the comments, modified for your repo. it's important that we *attach* the middle manager's prompt to our conversation vs. typing it out, so it never gets compacted. the middle manager's context window is holy. so our prompt tells it to spin up subagents — instead of doing things itself — for pretty much everything. 8. the prompt has some other cool, non-obvious details. firstly, every agent is told to keep working until it's met the burden of proof. in our case that's a (self)-review, bug bot being green, bug bot fixes being elegant, and a recording. we tell our middle manager to enforce this. secondly, we ask the middle manager to make sure linear stays up to date and communicate via the slack mcp. specifically, it should tag coworkers in slack whenever it has a question and post heartbeat updates in the eng channel. 9. your software factory is ready to go! in our case, it worked for 12+ hours without getting blocked. and produced extremely high quality prs. addendum: all of this is possible because with @AnthropicAI fable, models are finally smart enough to orchestrate/navigate large sets of work and their dependencies. and fable is a great software architect! now of course the 'downside' is that you're now review bottlenecked. we have 60+ PRs to review! but i think as an industry we should be able to come up with some innovation there. we'll certainly try our best to help here with @mainframe too! hope this is helpful! happy to answer any qs.
English
@petergyang 40+?
you just need a middle manager: x.com/sawyerhood/sta…
Sawyer Hood@sawyerhood
English
How many Codex or Claude Code threads can you pay attention to at once? Be honest 👇
English
if any company wants to sponsor my ability to side project maxx on vacation by giving me some cloud agent credits...
need a cloud agent with a vm to make progress from the streets of europe!
English
@pwnies @ListenLabs i think the success of bugbot shows there's a big generator-verifier gap with everything agents do, including this!
i'm not sure about your agents but mine make terrible product choices. especially when it comes to any "systems thinking" and reusing vs. introducing new concepts
English
@vinvan @ListenLabs A devil’s advocate argument against this: this would be amazing in a human era where people made common mistakes. Now with LLM development most agents will test this as part of the review flow. I personally wouldn’t spend the tokens checking twice, at least until tokens get cheap
English
it’s wild that no one has made a “product design bug bot” for code review:
- could a fresh computer-use agent (“user”) navigate your flow?
- is there an existing mental model in your product that you should’ve reused?
- would your users want it? (automated study via @ListenLabs)
English
English
@vinvan @dwarkesh_sp @3blue1brown this reminds me, i have a long-standing draft in my backlog titled 'is design np-complete' ... i want to finish it but i fear the audience is like 2 people
English
if you squint, this is the best podcast on ai and design that's come out this year: youtube.com/watch?v=TfyPsh…
the open problems in ai for math (e.g. "creativity") are similar to what we need to solve for better ai design/software.
wonderful listen @dwarkesh_sp @3blue1brown
YouTube
English
@seflless @ListenLabs looking at the website (and having read the blog post when it came out) it feels very visual design heavy.
whereas i'd imagine more emphasis on capital d design. how it fits in with your existing product, strategy, etc.
English
@jsngr @elirousso yes this is the most similar thing to it out there today!
would be awesome to see it go even deeper on the product design front. it's cool that it's already looking at the ux by auditing loading states etc!
English
@nateberkopec agree!
working on some of the harder parts right now. hope to make good progress there too.
English
@vinvan That’s what I’m saying, you haven’t done the hard part yet. Everyone is sitting on top of stacks of PRs they haven’t merged since Opus 4.5 came out. Plausible code generation is not a difficult solve.
English
If you have to review the code, you have a slop factory, not a software factory
Vincent van der Meulen@vinvan
fable produced 60+ (!) ready to merge prs for @mainframe overnight. here's how you can set up a software factory using fable as an orchestrator. prompt included! 1. narrate your *entire* todo list — everything that's top of mind — using dictation. i used @usemonologue and my prompt was ~10K words! you have to use voice because we want excruciating levels of detail. 2. locally, tell fable to plan all the work and put it in @linear. the goal is to make it such that a dumb model dropped in linear knows *exactly* what to do. aka all issues should have a clear DAG, plan, and optimize for max. parallelizability. i recommend running fable on high and having it fan out to codex subagents on xhigh (using the cc codex plugin.) to save usage. just make sure that fable itself is doing the architecture for the hardest issues! 3. now switch to @DevinAI. we'll use devin to run our factory because our orchestrator needs to be a cloud agent (it's going to run for a long time!), and be able to spin up other cloud subagents. 4. hook up the @linear @SlackHQ mcps. this will let your agents read the fable planned work, and communicate with you/your teammates. e.g. it can talk to your coworkers to make sure it's not stepping on anyone's feet! 5. a good feedback loop is *crucial.* before starting, make sure you have a bug bot (we use @cursor_ai.) and that your agents can test all its changes. e.g. we use the fantastic @limrun to make sure our agents can do ios development e2e. 6. create a devin ultra (fable) agent — this will be your orchestrator or 'middle manager.' we're going to tell it to look at linear and dispatch subagents for all tasks. the middle manager's goal is to keep spinning up subagents with the right level of intelligence until all the work is done. and keep all agents unblocked while you're sleeping, hanging out in the park, whatever! the reason the middle manager is able to orchestrate all of this is because fable planned everything in linear. and because the middle manager itself is also fable (contrary to most of its subagents.) 7. to prompt your middle manager, attach the middle manager prompt i'll post in the comments, modified for your repo. it's important that we *attach* the middle manager's prompt to our conversation vs. typing it out, so it never gets compacted. the middle manager's context window is holy. so our prompt tells it to spin up subagents — instead of doing things itself — for pretty much everything. 8. the prompt has some other cool, non-obvious details. firstly, every agent is told to keep working until it's met the burden of proof. in our case that's a (self)-review, bug bot being green, bug bot fixes being elegant, and a recording. we tell our middle manager to enforce this. secondly, we ask the middle manager to make sure linear stays up to date and communicate via the slack mcp. specifically, it should tag coworkers in slack whenever it has a question and post heartbeat updates in the eng channel. 9. your software factory is ready to go! in our case, it worked for 12+ hours without getting blocked. and produced extremely high quality prs. addendum: all of this is possible because with @AnthropicAI fable, models are finally smart enough to orchestrate/navigate large sets of work and their dependencies. and fable is a great software architect! now of course the 'downside' is that you're now review bottlenecked. we have 60+ PRs to review! but i think as an industry we should be able to come up with some innovation there. we'll certainly try our best to help here with @mainframe too! hope this is helpful! happy to answer any qs.
English
@rbright also any insights from 8 months ago are outdated. all models are so much better since dec 2025, as i'm sure you're aware.
would encourage you to at least try the experiment again!
English
@rbright i like to think it's a bit more than "shitting out PRs." there's still plenty of human handholding to prevent it from descending into madness, which shows in the quality of the prs: x.com/vinvan/status/…
and yeah — working to solve the verification problem as we speak!
Vincent van der Meulen@vinvan
agree that many models still struggle with software design! but i don't think what you're saying is true in this case: - fable (imo) can architect systems decently well. especially when it can mimic human created patterns, of which we have plenty - nowhere did i say that this is a dark factory! i'm a software engineer — i know and understand our architecture. my 10K words prompt isn't just "go do this." it includes ample notes on how i'd approach each task and what my hunches are - you really don't need world class software design for the average (frontend) PR. especially if you have a good AGENTS.md/skills that get models to avoid anti-patterns like effects
English
Yeah, I experimented with this like 8 months ago. It's a great way to end up with a bunch of incomprehensible slop that overwhelms whomever has to review it.
Shitting out PRs isn't hard. Can we stop pretending this is innovative while burning out teams of people with what is effectively a cheap party trick?
Instead of going "me too" on slop gen, solve the verification problems.
Vincent van der Meulen@vinvan
fable produced 60+ (!) ready to merge prs for @mainframe overnight. here's how you can set up a software factory using fable as an orchestrator. prompt included! 1. narrate your *entire* todo list — everything that's top of mind — using dictation. i used @usemonologue and my prompt was ~10K words! you have to use voice because we want excruciating levels of detail. 2. locally, tell fable to plan all the work and put it in @linear. the goal is to make it such that a dumb model dropped in linear knows *exactly* what to do. aka all issues should have a clear DAG, plan, and optimize for max. parallelizability. i recommend running fable on high and having it fan out to codex subagents on xhigh (using the cc codex plugin.) to save usage. just make sure that fable itself is doing the architecture for the hardest issues! 3. now switch to @DevinAI. we'll use devin to run our factory because our orchestrator needs to be a cloud agent (it's going to run for a long time!), and be able to spin up other cloud subagents. 4. hook up the @linear @SlackHQ mcps. this will let your agents read the fable planned work, and communicate with you/your teammates. e.g. it can talk to your coworkers to make sure it's not stepping on anyone's feet! 5. a good feedback loop is *crucial.* before starting, make sure you have a bug bot (we use @cursor_ai.) and that your agents can test all its changes. e.g. we use the fantastic @limrun to make sure our agents can do ios development e2e. 6. create a devin ultra (fable) agent — this will be your orchestrator or 'middle manager.' we're going to tell it to look at linear and dispatch subagents for all tasks. the middle manager's goal is to keep spinning up subagents with the right level of intelligence until all the work is done. and keep all agents unblocked while you're sleeping, hanging out in the park, whatever! the reason the middle manager is able to orchestrate all of this is because fable planned everything in linear. and because the middle manager itself is also fable (contrary to most of its subagents.) 7. to prompt your middle manager, attach the middle manager prompt i'll post in the comments, modified for your repo. it's important that we *attach* the middle manager's prompt to our conversation vs. typing it out, so it never gets compacted. the middle manager's context window is holy. so our prompt tells it to spin up subagents — instead of doing things itself — for pretty much everything. 8. the prompt has some other cool, non-obvious details. firstly, every agent is told to keep working until it's met the burden of proof. in our case that's a (self)-review, bug bot being green, bug bot fixes being elegant, and a recording. we tell our middle manager to enforce this. secondly, we ask the middle manager to make sure linear stays up to date and communicate via the slack mcp. specifically, it should tag coworkers in slack whenever it has a question and post heartbeat updates in the eng channel. 9. your software factory is ready to go! in our case, it worked for 12+ hours without getting blocked. and produced extremely high quality prs. addendum: all of this is possible because with @AnthropicAI fable, models are finally smart enough to orchestrate/navigate large sets of work and their dependencies. and fable is a great software architect! now of course the 'downside' is that you're now review bottlenecked. we have 60+ PRs to review! but i think as an industry we should be able to come up with some innovation there. we'll certainly try our best to help here with @mainframe too! hope this is helpful! happy to answer any qs.
English