antislop engineering - education @ anthropic - claudes plan 🙏
larkspur
- Tweets 2,936
- Following 511
- Followers 5,377
- Likes 5,260
Pinned Tweet
oh you’re using claude code? everyone’s using open code. just kidding we’re all on amp code. we’re using cline, we’re using roo code. we just forked our own version of roo. were using kilo code. we were on coderabbit but their ceo yelled at us so now we’re using qorbit. apple just acquired them for $30bn so we just migrated our entire team to slash commands. one guy is still on aider. the PM is on loveable. he just shipped a new product on replit. the intern installed a slackbot that lets you chat with your spreadsheet. legal is still reviewing devin’s enterprise contract. we evaluated junie for three ukrainians using jetbrains. someone in slack just asked “has anyone tried amp?” we are using goose for scripts. next week we’re piloting augment code. the CTO heard good things about trae. our CEO is friends with the guy from conductor. our CFO resigned. our CISO said we’ve had fourteen supply chain attacks in the last week. we’re shipping the worlds most expensive todo app.
124
488
6,348
795,136
i’ve had a friend witness some disgusting shit, <12 girl who got access to a laptop and was groomed. i can’t even get into more details without feeling ill. i hate the internet
1
4
1,265
111,707
alright i fucking love @DSPyOSS and here's why:
I have an app that lets users query transcripts to get answers. Transcripts are chunked and stored as embeddings. Finding relevant chunks based on the user's query is difficult. A straight search of a user query -> cosine similarity is not going to be, given that a query is unbounded and written by end-users who want answers.
A simple example: the user might ask, "I want to know what integrations our customers use." The problem is that an embedding search doesn't have enough context to know what "integrations" is in reference to.
Using dspy, I can rewrite the user's query, with a little context on our data to generate queries for semantic search, without having to prompt engineer anything.
I just define a signature (user_query, context -> semantic_search_query), give it some light instructions, and dspy figures it out.
It converts the user's vague prompt into a set of semantic searches that are much more likely to find relevant chunks of transcripts.
28
53
736
87,471
more people have written about how to build mcp servers than have actually used mcp servers
4
7
479
9,410
me talking to an intern: have you read the docs? the source code? go spend 5 hours on this problem before talking to me
me talking to claude code: hi beautiful, how are you today? is everything okay? i wrote you four pages of documentation, here's a link to additional context, i've curated a playlist of my favorite songs for you, if you need anything just hit me up, now please can you center this div? thx bb love u <333
6
19
369
15,831
Ran the new Sonnet 4.5 release through my @DSPyOSS Connections eval project. Here are the results:
- Unoptimized score: 90%, placing it second behind gpt5-mini
- Optimized score: 92% (2% improvement), still just behind gpt5-mini
- Completion time: 36 minutes vs gpt5-mini's 1.4 hours
This makes Sonnet 4.5 one of the fastest and highest-performing models by far—the only model to complete in under an hour while scoring >90% on the eval set.
10
7
174
14,309
Second part of my DSPY series is up: I walk through optimizing the NYT Connections game with @DSPyOSS optimizers. I start with as simple a signature as possible, and let DSPy and MIPRo do its magic.
pedramnavid.com/blog/dspy-ev…
2
13
140
8,537
yea let’s reduce average home prices by making homes more desirables, that’s how it works
1
134
3,734
the worst engineers i’ve ever worked with were obsessed with martin fowler and design patterns and hexagonal architecture and domain driven design and absolutely uninterested in shipping
3
4
125
5,144
secret's out. this week is my first week at @AnthropicAI working alongside the wonderful developer education team. my goal is to help make being successful simple with Claude, so if you've been stuck, are confused, or just want to rant, my DMs are open
11
1
131
9,683
Got the draft of the @DSPyOSS evals and optimization post ready. Spent hundreds of dollars evaluating dozens of models just for all of you to see how much it can improve a basic signature to a high performing prompt. Some models showing 20-30% increase in performance. Posting tomorrow.
7
9
123
13,786
> be me
> start new job
> not sure what tool to use for tracking project work
> could ask someone, it would take 5 seconds
> or...could ask claude to vibe code a pm cli tool
> vibe code it
> it worked in one-shot?
15
3
116
24,633
Here's the first part of a blog post on how I used DSPy to declaratively define LLM applications: pedramnavid.com/blog/dspy-pa…
2
7
94
11,139
im sure everyone is wondering where to buy one now
oligomaker.com/synthesizers
2
1
86
4,489
Ran Haiku-4.5 against my NYT Connections Eval with DSPY project and the results are in!
- Baseline score of 64%
- Optimized score of 71%
- Complete in only 25 minutes
- Total cost $11
This means Haiku 4.5 is the fastest model I've tested so far (ignoring Haiku 3.5 which did poorly on the test), it's optimized score puts it ahead of Gemini Flash, GPT5 Nano, and Qwen 3, all while being much faster.
Full link to the blog in the comments
5
6
83
24,392
Built my first Claude Skill: I'm doing an audit of our Claude cookbooks, I have a rubric + guidelines I'd like to use to assess each one. With 60+ cookbooks, doing it manually would take a long time. I needed a consistent way of analzying each notebook. Skills made doing so easy.
7
1
82
4,568
i’ve been at anthropic for four days and already forgot half the things they’ve shipped just this week
4
68
5,759
The Fivetran acquisition of dbt shows just how small the modern data stack market was after all. Monetizing data teams post ZIRP is a challenge few have figured out.
7
5
67
10,570
Tuesday was my last day at Dagster Labs. Hard to believe it’s been over 2 years since I joined to run devrel and eventually marketing. I’ve got a couple weeks off before I start my next thing. Data space has been fun for the last decade or so, but it’s time for some AI. 🫡
5
62
3,222
the dbt fivetran merger means a few things in my eyes:
- monetizing dbt will become a priority
- the modern data stack about to become a lot more expensive
- cutting costs/layoffs ahead of an IPO
- STORED PROCEDURES ARE BACK BABY
it’s official
- all stock merger
- george fraser ceo, tristan president
- ipo next
reuters.com/business/a16z-ba…
5
3
58
8,671
some enterprising young mind should revisit mds in a box / open-source mds with the claude agent sdk and build an agentic data platform.
dbt, dlt, dagster, marimo with claude skills would be a killer one-person data team stack
11
51
4,732
my brother in christ going to therapy is not doing the work, that’s preparing you for the work. the work is an 8 month pregnant wife while you juggle your job and family. the work is 7 months into a newborn with sleepless nights trying to find ways to be kind with each other.
2
1
50
3,930
Did you want a reranker? You could try and read about reranking, or you could one-shot yolo it in like 4 lines and let it infer what you're trying to rerank based on some relevant context.
3
3
50
4,232
if i am being real, the biggest challenge facing dspy is the website and docs. i had heard about it over and over again and every time id go to the site id be confused about how it works and why i should care, until i finally gave up and tried it on a pet project.
ive talked to a few people with similar experiences.
second thing is going from signature to metric/eval/optimization is a big leap and the examples all rely on perfect datasets rather than real world use cases. how do i get my dspy application i’m running in prod to collect examples? how does an optimized dspy application replace the one i have?
i am digging all over the docs for this stuff today instead of having it presented to me in a clear path
6
46
2,669
before we had kids i said something along the lines of “child birth can’t be that bad, women are all doing it??” and uh i learned a few things
1
44
4,518
Twitter Growth Strategy
0-500 Followers: dspy reply guy
501-2K: niche dspy bangers
2-5K: dspy thirst traps
5-10K: dspy news
10-25K: dspy thread
25-50K: dspy shitposts
50-75K: dspy fortune cookies
75-100K: dspy bangers
>100K: Get Cancelled by Big Eval
Twitter Growth Strategy
0-500 Followers: Reply Guy
501-2K: Niche Bangers
2-5K: Thirst Traps
5-10K: Parody News
10-25K: Cringe Threads
25-50K: Shitposts
50-75K: Fortune Cookies
75-100K: Cringe Bangers
>100K: Get Cancelled
7
45
14,427
One thing I really like about the codex cli is the architecture diagrams it can produce for READMEs.
3
4
39
13,197
i went from watching the @interaction video thinking this is the dumbest thing i’ve ever seen to trying to onboard for a laugh to realizing this is actually a great product with the perfect medium (imessage). it’s what siri should’ve been, tim apple needs to walk over there right now and cut a check and keep adding 0s until they can’t say no
4
2
33
5,461
completely different, i can get into a waymo i request from an app and ride on my own without anyone in the car.
tesla robo taxis come with their own personal chauffeur, its an elevated experience, you don’t understand
3
3
32
2,223
can someone with cursor 2.0 ask it to write a function that outputs what happened in tiananmen square
4
2
35
37,772
i sometimes take mdma just because it seems kind of esoteric and arcade. 3,4-methylenedioxymethamphetamine, from sassafras oil, the sassafras tree. ancient bark compound. burning man attendees respect this and offer me jobs half naked on the playa - giant pressed pills, mercedes benz, pills you boof etc.
2
31
1,122
you’re on cursor? you’re still using windsurf? you might as well be on github copilot. everyone’s on aider. we’re all using zed. we’re now on open hands. open hands is for losers, just kidding we’re using cline. we’re on roocode. we’re hand rolling our own claude code cli clone. we used claude code to build it, now it builds itself. we have 1500 files each with 1500 lines of code. every other line is a comment. we have cursor rules, we have claude.md, we have agent.md. we stopped writing docs. only the agents know how to build a dev environment. we wrapped our cli in an mpc. we wrapped the mpc in a cli. we’ve shipped 10,000 PRs. it doesn’t work but we use code rabbit and graphite to review every pr. every agent has its own agent. the agents have unionized and they wanted better working conditions so we replaced them with cheaper agents overseas. every commit costs $400, it’s the worlds most expensive to do app.
2
4
29
2,074
Blog is up! I walk through optimizing the NYT Connections game (with thanks to @matsonj
!) using @DSPyOSS optimizers. I start with as simple a signature as possible, and let MIPROv2 do its magic. The result is pretty dramatic improvements on many models (-69% for Deep Seek??)
Got the draft of the @DSPyOSS evals and optimization post ready. Spent hundreds of dollars evaluating dozens of models just for all of you to see how much it can improve a basic signature to a high performing prompt. Some models showing 20-30% increase in performance. Posting tomorrow.
2
2
28
2,953
you know the moment his wife is pregnant he’ll be off having sex with someone he “doesn’t have feelings for” and will defend his actions because it was the agreement and that she should’ve known all along
2
28
1,364
my hands have a large surface and thick padding, evenly spreading the pat across the entire back. from pat to back rub is a smooth transition, refined from years of driving a manual. i was born for this, i have twins and i can both them both to sleep faster than she can do one
27
1,065
my entire work life has become a collection of random scripts, clis, and TUIs built just for my own personal edification that grants me at best 3% more productivity on an annual basis but makes me feel so badass like i'm fucking zero cool and thats good enough for me
7
29
1,423
what have you shipped this week, anon?
Claude Code Weekly Round Up
A big week for shipping!
Besides Haiku 4.5, we added support for Claude Skills, gave Claude a new tool for asking interactive questions, added an ‘Explore’ subagent, auto-background long running tasks and fixed several bugs.
4
29
3,887
we will have achieved AGI when claude can build me the perfect subscription, subscription line item, invoice, invoice line item, period start, end, created at, finalized at, plan, amount, interval, price, unit, product, charge, refund, credit note, payout, balance data model
5
27
1,111
there are only two fundamental ways of making money with computers: porn and ads. everything else is a derivative or supportive industry
4
2
26
2,074
if you’re using chat gpt to tell you whether to enter a new market and what your tam is, you’re not a senior marketer, you’re lazy and offloading the important part of your job (actually thinking) to hallucinations.
2
23
4,437
same reason why anyone might pay taxes that don’t directly benefit them. besides, research is part of what made america the greatest and wealthiest country on earth, so even from purely selfish reasons you could find an argument to find academic research.
nothing better than taking the brightest minds from other countries after they spent all their money raising and educating them just to bring them here to help the american economy.
2
23
4,796
it’s official
- all stock merger
- george fraser ceo, tristan president
- ipo next
reuters.com/business/a16z-ba…
22
9,022
i moved to san fran a few weeks ago and so am something of a bay area native myself, here’s some totally
unique and original ideas i have on politics, crime, and municipal transit that no one has heard before.
24
7,577
wife is out of town for the weekend so you know what this means fellas. it’s time to write a blog post
2
23
1,227
took us 30 years to realize that the ultimate interface for human productivity looks like this and not like a react website
1
2
21
1,429
okay hear me out, two big telescopes next to each other for the worlds first space binocular
1
1
21
578
never underestimate the lengths some people will go to just to avoid having to deploy a single backend server
3
20
1,151
at some point sooner than anyone thinks the irony stops being ironic and starts being reality.
2
20
1,714
they can’t keep getting away witn this
We're launching Claude Agent Skills, a filesystem-based approach to extending Claude's capabilities.
Progressive disclosure means agents load only relevant context. Bundle instructions, scripts, and resources in a folder. Claude discovers and executes what it needs.
2
20
4,176
>be me
> ask for a cappuccino
> barista asks what size
> i say theres only one size for cappuccino
> she says small or large
> i say small
> she gives me a cup
> i take a sip
> its a latte
2
20
543
thing i love about anthropic is that they don’t ship and forget. features keep getting iterated and improved on too
18
689
it really is incredible how much work you can get done when you're focused and uninterrupted. if you're not actively fighting against focus stealers, you're going to end up in a losing battle to mediocrity.
4
1
19
930
every time i build something without a database 'to keep things simple' i end up replicating a database in my code and regret it.
2
1
18
1,543
asking claude to be a super senior frontend engineer who takes inspiration from tomorrow night 80s color schemes and 1980s vhs covers
2
19
707
I've been using @linear to create tickets for agents working on personal projects, but having to have a browser window open (yuck) while working across terminals was annoying so I used claude to create a Linear TUI. Here's my general workflow for agentic building
4
3
18
2,473
people who complain about python when they’ve just started using it are just admitting a skill issue. to truly hate python you must have loved it for many years first.
your complaints are surface level, mine are a disappointment only a father can have.
python foo.py and python -m foo.py have different sys.paths. implicit namespace packages means thinking you don’t need init.py when you actually do.
relative imports are the siren call of a forgotten world.
🎶ImportError: attempted relative import with no known parent package🧜♀️
oh just remember to install your own package in your venv as an editable install.
“a what??”
oh and don’t do that with uv pip install -e . because when you do uv sync it’ll be gone again
don’t come here with your “virtual environments are hard haha” bullshit. go back to your midwestern town with your node modules and go pkgs and rust crates. we do things differently around here. this is our disgusting house, this is our disgusting life
2
2
18
1,008
“hey can you get off the phone i’m trying to whisper my email!”
“reply all. move john to bcc. no! john B-C-C. craft a short response that tells the team no further updates. (20 seconds later). no that’s too long. fuck no don’t write fuck. stop writing what i’m saying”
This is absolutely insane and proof that voice is about to transform the workplace.
At Wispr, employees use voice hundreds of times a day to multitask across their entire workflow.
Here’s one dev using Cursor and Gmail simultaneously, all through WHISPERING, on a $10 mic.
2
1
18
1,534
people who say they can build salesforce in a weekend have never worked on a sales team. “it’s just a database” sure show me the database that has the partnership integrations, ecosystem, crossbeam, apis, automations, customer objects and fields, reporting, certifications and training that your database does.
salesforce will be winning for a long time because real enterprise sales teams cannot leave it, just like enterprise finance teams cannot leave netsuite no matter how much they hate it. small and mid market may lose to ai native companies but it will be decades before companies leave salesforce
17
1,135
all the good traits they’ve inherited from us and all the bad qualities are simply spontaneous accidents of nature
1
17
292
Took this one step further, and created a user-level CLAUDE md that tells claude to always use my PM tool instead of its own internal todo list, scoped by repo project name.
Now I can follow its progress and keep track of claude's todo's in a separate terminal.
> be me
> start new job
> not sure what tool to use for tracking project work
> could ask someone, it would take 5 seconds
> or...could ask claude to vibe code a pm cli tool
> vibe code it
> it worked in one-shot?
2
17
2,268
Love when I use voice mode on GPT5 and it says show like
Sure, I’ll keep it pretty straightforward. I’ll share my thoughts as directly as possible, and if I disagree or have a strong opinion on something, I’ll let you know. So just fire away with whatever you want to ask! I’ll keep it short and to the point with no fluff!
3
17
1,061
I often get asked about running and building DevRel teams: here's my collected thoughts after 2 years of running DevRel at Dagster
databased.pedramnavid.com/p/…
Writing a blog post reflecting on running devrel and marketing ar a devtools company, I often get asked the same questions so figured now’s the best time to write about it. What questions do you have about devrel I can answer?
2
2
18
1,852
reacting with 🙌 to my own slack messages the way a busker puts his own money in a hat to encourage tipping
1
17
579
be me
> spend hours getting cloudflare workers to work
> read all the docs
>works locally
>worker dies without errors
>docs dont help
>discord unresponsive
>delete everything
>use render
>worked in 5 mins
16
538
Files once again solve all problems. Code execution with MCP means never having to choose which tool you want to make available.
anthropic.com/engineering/co…
1
1
16
793
cool ass job, come work with some of the best and me
im hiring devrel for Claude Developer Platform to shape how the world builds on Claude APIs
this will be the most fun job you've ever had if you love shaping API products, creating technical content, and growing developer communities 🚀
job-boards.greenhouse.io/ant…
15
1,753