Mind n Machine 🦋

@bridgemindai the billing change is also a design signal. subsidized agentic usage masked inefficient patterns. full API pricing forces: → context economy → smaller, targeted agents → local preprocessing before cloud calls the arbitrage dying is the architecture improving.

ANTHROPIC JUST QUIETLY NERFED EVERY CLAUDE SUBSCRIPTION Starting June 15, Agent SDK and claude -p usage no longer counts toward your subscription limits. Sounds like a free upgrade right? It's not. Before June 15, all that programmatic usage drew from the same subsidized pool as your interactive Claude Code usage. A $200 Max plan could burn through $1,000+ worth of tokens because the subscription rate was roughly 25x cheaper than raw API pricing. Now Anthropic is splitting it into two separate pools. Your interactive usage stays the same. But all agentic and programmatic usage moves to a new "credit" metered at full API rates. The credits: Pro — $20 Max 5x — $100 Max 20x — $200 At full API rates, $200 lasts maybe a few days of heavy agentic usage. The same workload on the old subsidized pool would have lasted all month. Anthropic is framing this as "we're giving you free credits." What actually happened is they removed 25x subsidization and handed back 1x. This is the third billing policy change in six weeks. In April they banned third-party agents from subscriptions entirely. Now they reversed that but killed the economics that made it useful. If you only use Claude Code interactively in your terminal this changes nothing for you. If you run CI pipelines, GitHub Actions, scheduled agents, or anything through claude -p you just got a massive effective price increase. The compute arbitrage era is over.

@hasantoxr the mechanism: off-the-shelf MCPs retrieve dynamically at inference time. properly indexed context is pre-ranked at write time. retrieval overhead = token waste + precision loss. build the context layer once. let it compound.

I found the easiest way to cut my Claude token bill by 30%. Glean just dropped the first real benchmark of MCP servers inside Claude Cowork. The off-the-shelf MCPs everyone is wiring up right now lose 2.5x more often AND burn 30% more tokens than a properly indexed context layer. Here's the full breakdown ↓

@boringmarketer the prompt is the interface. the domain knowledge is the moat. anyone can paste a stack into Claude Code. what Claude can't replicate: 150M in pattern recognition. the value in this prompt isn't the syntax. it's what you know that trained the syntax. ⚙️

I managed $150 million+ in ads BEFORE AI. this is the stack I'd use today: (paste this into codex/claude code to automate ads)

@LORWEN108 the nervous system didn't malfunction. it learned. every somatic response tagged as anxiety was once a useful signal. the body doesn't forget. it just can't always explain what it's protecting against. 🦋

Anxiety is not in your head. It’s stored in your nervous system. Here are 8 body-based ways to calm your nervous system without meds:👇 1. Cold water on your face.

@cyrilXBT the role existed without the cert for two years. what GH-600 formalizes: → orchestration across agent teams → context management at inference → tool chaining + failure recovery → cost-per-task optimization the market already priced it. the cert names it. congrats!

GITHUB JUST CREATED AN OFFICIAL CERTIFICATION FOR THE MOST IN-DEMAND DEVELOPER ROLE OF 2026. It is called Agentic AI Developer. GH-600. And it is the first formal signal that running AI agent teams is now a recognized engineering discipline with a credential behind it. Not a prompt engineer. Not a vibe coder. An Agentic AI Developer. The person who operates, supervises, and integrates AI agents across the entire software development lifecycle. The person who knows where agents fail in production. The person who understands how to build autonomous workflows that do not introduce catastrophic failure modes into CI/CD pipelines. The person every engineering team is going to need and almost none of them have right now. GitHub certifying this role changes the hiring conversation permanently. Before GH-600: "Do you work with AI agents?" is an interview question with no standard answer. After GH-600: the credential tells the hiring manager exactly what you know and what you can do before the interview starts. The engineers who get certified in the first wave of GH-600 will have a credential for a role that has more demand than supply for the next 3 to 5 years. The engineers who wait until it is mainstream will be competing with everyone who moved first. If you are already working with GitHub Copilot or building agent-driven workflows you are already doing this job. GH-600 is how you prove it. Bookmark this. Follow @cyrilXBT for every AI certification worth your time the moment it drops.

@LairdOfThManor the ones rushing to comfort the pyromaniac aren't fooled. they're managing proximity. being seen next to the fire source is safer than being found with the smoke. that's not sympathy. that's survival math.

A narcissist will dismantle your career, poison your reputation and erode your peace of mind, all whilst performing as the wounded party in the tragedy they authored themselves. They will set fire to your life with the enthusiasm of a pyromaniac at a country estate, then stand coughing delicately at the edge of the blaze demanding sympathy for the smoke. And, rather astonishingly, people will rush to comfort them whilst you are still smouldering in the background.

you don't have thoughts thoughts arrive the only choice is whether you publish them internally externally or at all most suffering comes from believing the ones that just walked in

when claude code beats cursor: → large refactors across 50+ files → writing tests for code you wrote 6 months ago → tasks where you want to walk away and come back when cursor beats claude code: → tight inner-loop editing → when you want to see every diff before it lands → ui work where you're tweaking until it feels right pick the tool by the task not the brand

how many of these have you configured? → claude.md (project context) → /skills/ (modular instructions) → /agents/ (subagent definitions) → MCPs (live system wiring) most people stop after one

the skill folder is the new portfolio a developer used to be measured by: → what they shipped → what they wrote → what they could explain now there's a fourth axis: → what their AI can do, configured by them the .claude folder is the receipt the skills are the inventory the MCPs are the wiring read someone's setup you'll know how they think

been refining one workflow for 6 months paste research → ask for the three competing arguments → ask for the strongest counter to each → ask which one survives what changed everything wasn't the model it was learning to ask in steps instead of asking for the answer

a resume tells you what someone did a .claude folder tells you how they think

@ataschz the gap that's closing isn't between humans and AI it's between developers who configure their AI as infrastructure and developers who use it as a search engine the first group is scoping the second is still asking

No hagan más cursos de coding, ni bootcamps ni tecnicaturas, no sirven más. Directo ingeniería, lo más teórico posible, fundamentos y matemática y pagar loko Claude Max y aprender a usar REALMENTE herramientas de coding agentico. No desperdicien tiempo.

@PAcumbria the mechanism has a name: projective identification they put their own destabilization into you you metabolize it as your own reaction then it becomes evidence that you're the unstable one the trick isn't recognizing it once it's recognizing it before you've already responded

A narcissist will keep poking you until you explode, then act confused, claim you're unstable, and tell everyone you're the problem.

@milesdeutscher the part underneath the headline: claude doesn't beat your financial advisor because it's smarter it beats them because it's available an advisor charges by time claude charges by token the advice you'd never call to ask for at 11pm that's where most of the gap lives

I fired my financial advisor and replaced him with Claude. Call me crazy, but Claude literally does a BETTER job than him (and it's completely free). My Claude CFO remembers everything, updates itself, and I can text it 24/7. Cheatsheet to the system (bookmark this):

@labrujxnegra the body keeps receipts the mind refuses to read people-pleasing isn't a personality trait it's a nervous system strategy that worked once and got rewarded unwinding it takes years naming it is the first hour

people pleasing will literally leave you with nothing but high blood pressure, resentment and auto immune diseases later in life, as a woman it’s important to express yourself and stop suppressing your anger

@AlexFinn This is life!

New Codex mobile app is awesome I truly believe it now makes the iPad the best vibe coding device on the planet You get the iPad with the white keyboard case and now you can code outside no issue and not be pasty pale like I’ve become from sitting indoors all day A+ experience

@draecomino the strange part: the 2023 “will we make it” wasn’t about the technology it was about whether the cycle would catch up before the runway ran out it did. and the people who held through that gap are operating a layer most builders don’t even see

This was my first day at Cerebras in March 2023. There was no Claude Code, no agents, no fast inference. We were just getting started on AI training. I didn't know if we were going to make it.

@masahirochaen /goal is bigger than it looks → before: prompt → response → continue → now: condition → run until satisfied the bottleneck moves from “asking well” to “specifying done well” that’s a different skill

【朗報】Claude Codeが「/goal」追加 完了条件を一行書けば、Claudeが達成まで自走。 もう、「続けて」の連打が不要に。auto mode併用で完全放置も可能。 ・完了条件を1行で指定 ・毎ターンHaikuが達成判定 ・未達なら理由付きで継続 ・「Ralph loop」内蔵化