标签：#ai

Stacking, Bidirectionality, the Encoder-Decoder, and LSTMs Last post ended with a simple RNN and three promises: LSTMs, bidirectional RNNs, and attention. This post delivers the first two, plus the refinements that turn a working-on-paper RNN into something you'd actually deploy. By the end, you'll know how to stack RNNs for depth, why reading a sentence backward as well as forward (bidirectionality) makes representations sharper, how the encoder-decoder turns one sequence into a different one for machine translation, and exactly what breaks in a simple RNN that LSTMs and GRUs were invented to fix. Attention, the fix for the last problem we'll hit, gets its own home in the transformer post, so we'll stop right at the edge of it. The simple RNN was the idea. This post is the engineering. A vanilla RNN carries a thread of hidden state through time, but in practice, that thread frays on long sequences, only sees the past, and bottlenecks everything through one final vector. Each section here is a fix for one of those problems. Put them together, and the path from "RNN" to "transformer" looks less like a leap and more like a series of obvious next steps. Stacking RNNs for Depth The first refinement is the easy one. Nothing says an RNN's output has to go straight to a prediction. You can feed the entire output sequence of one RNN as the input sequence to another. Then another. These are stacked RNNs (also called deep RNNs), and they usually outperform a single layer. Why does depth help? The same reason it helps in vision. Each layer learns representations at a different level of abstraction. The lower layers pick up fundamental, local properties; in language, that's roughly the level of parts of speech and named entities. The higher layers compose those into bigger groupings: descriptive phrases, "this is the answer to a question," and so on. We can't point at layer 3 and say "this one does coreference." But the theory holds up well enough that researchers have probed m

The Question Nobody Wants to Ask: Does Your Skill Actually Help? You spent an afternoon crafting a carefully structured Skill for your agent. Clear steps, thorough edge-case notes, well-formatted output requirements. You tested it manually a few times, the outputs looked great. You shipped it. Three weeks later, you notice that some task success rates have gone down compared to before the Skill existed. This is not a hypothetical. In May 2026, Microsoft Research published two concurrent papers — SkillLens ("From Raw Experience to Skill Consumption") and SkillOpt ("Executive Strategy for Self-Evolving Agent Skills") — that measured this failure mode at scale. Their finding: negative transfer happens in 25% of cases , and you cannot reliably identify the bad skills just by reading the text. One paper answers "why skills sometimes backfire." The other answers "how to make skills systematically better." Together they sketch a new paradigm for agent capability improvement. Part One: SkillLens — Mapping the Full Skill Lifecycle A Skill Is Not a Point — It's a Pipeline Most practitioners think of a Skill as "a block of text instructions for an agent." SkillLens decomposes this into a three-stage lifecycle : Stage 1: Experience Generation Target model M runs training tasks, producing an experience pool of trajectories (both successes and failures) ↓ Stage 2: Skill Extraction Extractor model E distills the experience pool into a structured skill document — procedural knowledge under a fixed budget ↓ Stage 3: Skill Consumption The same target model M, equipped with the extracted skill, is evaluated on held-out test tasks Notice there are two distinct roles in this chain: the Extractor (distills knowledge from trajectories) and the Target (consumes knowledge to improve task performance). SkillLens's central insight is that these two roles are independent — a strong task executor is not necessarily a strong extractor, and vice versa . Two New Metrics: EE and TE To separate thes

I used to deploy Node.js apps on EC2 and manage servers like it was my second job. Port configs. PM2 restarts. Nginx rewrites. SSL renewals. Then I ran my first AWS Lambda function. 80% of that work is gone. Here's what Lambda actually does that nobody explains clearly: → You write a function → AWS runs it ONLY when triggered → You pay for milliseconds of execution → It scales from 1 to 1,000,000 requests without you touching anything As a full-stack developer in Bahrain, preparing for my AWS Developer Associate exam, this is the shift that changes how you think about backend architecture. Not "how do I manage a server" but "what should happen when this event fires." That mental model switch took me a week to fully get. I'm documenting everything as I study. Drop a 🔥 if you want me to share my Lambda notes weekly.

By Takeshi Yokoyama — Onecarat Labs Hi. I'm Yokoyama, and I build a local-first AI text editor as a side project, along with a few other experimental tools. Working on them, I keep running into the same question about where the web is going. This post is one observation, plus a small experiment I built to test it — including a Chrome extension you can actually try. The short version: I think websites will increasingly be read through AI agents, reshaped per reader, on the fly. And once that happens, there's a clear gap between sites that are easy for an AI to read and sites that aren't. What's starting to happen Until now, people read websites as websites. You open the top page, follow the menu, read the body, click a button — tracing the path the maker designed. As local AI and AI agents become normal, that breaks. People stop opening the page directly. They tell an AI what they want — "Can I try this quickly?" , "I just want to check it's safe" , "Just the gist" — and the AI reads the web and reshapes it into the form that reader wants. What the reader receives is no longer the layout the maker built. This isn't speculation. The idea that AI generates the interface for the reader already has a name — Generative UI — and it's one of the hottest areas in frontend right now, with Google, Vercel and others building toward it. But notice who's holding the pen in almost every version of that story: the site , or an AI embedded in an app — something under the maker's control. What I'm looking at is one step past that: a local AI, in the reader's own hands, reshaping any site into that person's preferred form — with no involvement from the maker at all. The initiative moves from the maker to the reader. The part that nags at me as a builder I build software too. So this shift nags at me. A site carries its maker's intent and rights. The order things appear in, what gets emphasized, the tone. Design, copy, flow — all of it is deliberate. Having an AI quietly reorder, rewri

In the past, when people heard the word “token,” they usually thought about cryptocurrency, trading,...

Agent Memory: Short-Term, Long-Term, and Episodic Main Thumbnail Image Prompt: A human brain cross-section illustration in neon tones on dark background. Three regions clearly demarcated and labeled. The hippocampus region glows blue, labeled "Episodic Memory: what happened." The prefrontal cortex glows orange, labeled "Working Memory: what I'm doing now." A network of distributed nodes glows green, labeled "Semantic Memory: what I know." Arrows show information flowing between regions. Scientific but accessible, the memory architecture made neural and visual. Memory Architecture Diagram Image Prompt: Four storage boxes arranged vertically on dark background. Top: "In-Context Window (Working Memory)" — fastest, smallest, temporary, shown as RAM chip icon. Second: "External Vector Store (Semantic Memory)" — fast retrieval, persistent, shown as cylinder with search icon. Third: "Key-Value Store (Episodic Memory)" — structured facts, shown as database icon. Bottom: "Fine-Tuned Weights (Procedural Memory)" — slowest to update, most permanent, shown as brain with lock. Arrows showing read/write speeds between boxes. Clean, technical, the hierarchy is the insight. Memory Retrieval Flow Image Prompt: A query arrives at an agent on the left. Four parallel arrows go right to four memory sources: conversation history (short chat bubbles), vector database (semantic search visualization), structured database (table icon), model weights (brain icon). Each source returns relevant items. A "Memory Fusion" box on the right combines the results. The agent sees an enriched context. The retrieval from multiple stores is the architecture. Every conversation with an LLM starts from zero. You explain your project. You explain your preferences. You explain your constraints. You spend five minutes providing context. You come back tomorrow. You do it all again. The model remembers nothing between sessions. The context window closes. The state is gone. Every interaction is the agent's first

174 million smart TVs, most of which run Linux. 3.9 billion Android phones. Zero marketing. Tonight, somewhere around the world, a person will press the power button on their Samsung TV. A proprietary Samsung logo will appear. A polished menu will load. They will open Netflix, scroll through recommendations, and pick a movie. They will never know that every frame they see is being scheduled, managed, and rendered by a Linux kernel, the invisible engine that sits between apps and hardware. They will then reach for their Android phone to check something on social media. Another Linux kernel. If they are sitting in a Tesla, the touchscreen showing their charging status is running yet another Linux kernel. The “year of the Linux desktop” debate has been running for two decades. Entire forums exist to argue about whether 2025, 2026, or 2027 will finally be the year Linux takes over the PC market.

You've seen this before. You ask your AI agent: "Find ∫ x·e^x dx" It confidently replies: e^x + C , complete with a plausible-looking derivation. You nod. Then you check — the correct answer is (x−1)·e^x + C . It was wrong by a mile, and you almost shipped it. This is the fundamental problem with AI math today: LLMs can talk, but they can't verify their own work. They sound convincing while being catastrophically wrong. And the more complex the problem, the better the hallucination. Math.skill changes that. It's an open-source mathematical reasoning skill for AI agents — install it, and your agent stops guessing and starts verifying. What Makes It Different Typical AI Math Plugin Math.skill Workflow Prompt → LLM → answer Prompt → 7-step pipeline → ≥2 verifications → answer Verification None Answer blocked if verification fails Open problems Might hallucinate a "solution" Honestly says "this is unsolved" Error recovery No mechanism Auto-backtrack, fix, recompute, re-verify The core differentiator: a verification engine that runs at least 2 of 11 independent checks on every answer. No answer leaves the pipeline unverified. Period. The 7-Step Pipeline Every problem flows through this: Step What Happens Why It Matters 1. Parse Extract conditions, goals, variables, implicit domain constraints Catches misread problems before they waste your time 2. Model Build formal representation: equation, function, matrix, probability space, etc. Prevents building the wrong mathematical structure 3. Select Choose the optimal method from 30+ strategies Avoids brute-forcing when elegance exists 4. Solve Step-by-step with mathematical justification at every transformation Full traceability — nothing hidden 5. Verify Apply ≥2 of 11 independent verification methods The differentiator — catches what LLMs miss 6. Correct If verification fails: backtrack to last known-good step, fix, recompute, re-verify No "doubling down" on wrong answers 7. Deliver Exact answer (not approximate), domain con

AI applications are moving quickly from simple chatbots to systems that can search, reason, recommend, summarize, and act on live business data. For developers, that usually means wiring together databases, embedding models, vector search, rerankers, orchestration logic, and application code. For no-code AI builders, it often means waiting for those integrations to exist before an idea can become a working prototype. The MongoDB extensions for Dify help close that gap. With the new MongoDB Atlas and Voyage AI extensions, Dify builders can visually compose AI workflows and agents that connect directly to MongoDB data, perform semantic retrieval with Atlas Vector Search, improve result quality with Voyage AI embeddings and reranking, and optionally interact with operational documents through controlled database tools. The result is a practical path from idea to working AI application: less custom orchestration code, more reusable building blocks, and a smoother experience for both developers and no-code builders. Why Dify and MongoDB Belong Together Dify provides a visual environment for building AI apps, workflows, and agents. It makes it easy to connect user input, model calls, tools, prompts, and outputs into a working application. MongoDB Atlas provides the data foundation: flexible documents, operational queries, aggregation, full-text search, and vector search in one platform. Together, they create a powerful pattern: Dify orchestrates the AI experience — workflows, agents, prompts, tools, and user interactions. MongoDB Atlas stores and retrieves the data — documents, application records, knowledge sources, and vector embeddings. Voyage AI improves retrieval quality — embeddings for semantic search and reranking for precision. For a no-code builder, this means you can assemble a retrieval-augmented generation workflow visually. For a developer, it means the integration points are packaged as reusable Dify tools rather than one-off glue code. Meet the Extensions

Intel Foundry's Rio Rancho Facility Moves Toward Glass Substrate Volume Production Reports from Wccftech and Forbes (May 26, 2026) indicate that Intel Foundry's facility in Rio Rancho, New Mexico, is advancing toward becoming the world's first factory to achieve mass production of glass substrates — a next-generation chip packaging technology considered critical for scaling AI hardware beyond current organic substrate limitations. The facility has already begun manufacturing silicon photonics products for external customers and is expected to play a central role in Intel's advanced packaging strategy. Why Glass Substrates Matter for AI Glass substrates address fundamental limitations of current organic (ABF) substrates that are becoming bottlenecks for AI chip scaling: Extreme flatness (<1 μm warpage) enables larger die and chiplet assemblies Low CTE (3-8 ppm/°C) closely matches silicon (2.6 ppm/°C), reducing thermal stress Higher interconnect density due to dimensional stability Better high-frequency performance with low dielectric loss Larger format supporting bigger interposers than organic substrates For AI accelerators that already push CoWoS substrate limits at 5,500+ mm², glass substrates could enable even larger multi-chiplet assemblies. Intel's Advanced Packaging Ecosystem Intel has been building an advanced packaging portfolio: EMIB (Embedded Multi-die Interconnect Bridge): High-density die-to-die connections Foveros : 3D stacking for logic-on-logic packaging Co-Packaged Optics (CPO) : Recently demonstrated glass-core substrate prototypes with CPO Customer Base According to Forbes: Existing customers : AWS, Cisco Reportedly in discussion : Apple, Google, Microsoft, Nvidia, Tesla Commercial Timeline Milestone Timeline Glass substrate R&D announcement 2023 Pilot line (Chandler, AZ) 2024-2025 Silicon photonics production (Rio Rancho) 2026 (active) Glass substrate volume production ~2028-2030 Global Competition Intensifying SKC/Absolics (Korea): Operating pilo

You’re Overpaying Every Day — You Just Can’t See It Think about the last time you asked an AI to clean up your meeting notes. You probably opened a new chat, pasted in the transcript — maybe 1,500 words — then pasted your usual notes template on top of that, then said something like “format this, bold the action items.” It worked. Useful, even. But here’s what actually happened: the model just read ~3,000 words to produce ~300 words of output. Do that five times a week. Every week. And now think about what’s riding along in that context every single time — your template, your formatting preferences, all the background you’ve already explained before. The model doesn’t remember any of it. It reads it fresh on every call. Every repeat. Every charge. This isn’t a flaw in ChatGPT. It’s the fundamental nature of chat as a paradigm. 2. Chat Is Great — But It Has a Structural Bug Chat is the most natural way to start with AI. Unclear what you want? Talk it out. Need to change direction? Just say so. The feedback loop is instant, the barrier is zero. That’s why everyone starts there. But chat has a structural problem: every single turn carries the entire history. This is how context windows work — the “conversation history the model reads every single time.” Every API call packages up your full history and sends it to the model. You pay for every token the model reads. Ten rounds in, round ten doesn’t cost the price of one message. It costs the price of all ten, stacked. Here’s a concrete version of this. Say you use AI to write your weekly status update. You paste in your bullet points from the week, say “turn this into a proper update,” tweak the tone, go back and forth a couple times. Feels efficient. But those bullets, plus the AI’s draft, plus your follow-up messages, plus the format you’re implicitly re-explaining each time — the real token cost of one weekly update is probably 5 to 8x what you’d guess. You’re paying for repeated context. The bill just isn’t obvious e

🔮 Hermes Agent 🤖: A Practical Guide 🔥 — and How It Stacks Up Against OpenClaw & GoClaw 📊 Hermes Agent Challenge Submission Truong Phung Truong Phung Truong Phung Follow May 18 🔮 Hermes Agent 🤖: A Practical Guide 🔥 — and How It Stacks Up Against OpenClaw & GoClaw 📊 # hermesagentchallenge # devchallenge # agents 7 reactions Comments Add Comment 16 min read

How I Built a Bilingual AI-Powered Business Manager for Street Vendors in Jamshedpur The Problem Walk through any street in Jamshedpur, Jharkhand, and you'll find hundreds of small shop owners and hawkers running their entire business from memory — stock levels, daily sales, employee wages, profit margins. They have no tools. Notebooks get lost. Mental math fails. And at the end of the day, most don't even know if they made a profit. These vendors are not tech-illiterate. They have smartphones. They use WhatsApp. But every existing business app is either too complex, too expensive, or only available in English. I wanted to solve that. The Solution — Dukaan Manager Dukaan Manager is a free, offline-capable Progressive Web App (PWA) and Android app built for small shop owners and street vendors in India. It runs entirely in the browser — no server, no subscription, no data charges beyond the initial load. Built using HTML, CSS, JavaScript , and powered by the Gemini 2.0 Flash API , it gives every small vendor access to a smart business advisor in their pocket. What It Does 📦 Stock Management Vendors can add items by name or photo. Each item stores buying price, selling price, and quantity. The app automatically calculates profit margins and flags low-stock items with colour-coded alerts. 💰 Daily Sales Recording Sales are recorded item by item, linked to which employee made the sale. The app auto-fills prices and calculates the total. Unsaved drafts survive accidental page refreshes using sessionStorage. 👥 Employee Tracking Add employees with their role, salary, phone number, and joining date. Mark daily attendance (present/absent) with one tap. Track total sales made by each employee across all time. 📅 Sales History Every saved sale is stored permanently in the browser's localStorage. History is grouped by date, collapsible, and shows daily revenue and profit — clean and simple. 🤖 AI Business Advisor (Gemini-Powered) This is where Google AI comes in. Using the Gemini

The interesting result isn't who won. It's that across four security domains, Claude and Gemini missed the same hardening steps — and if you've shipped AI-generated auth middleware this year, your code almost certainly has the same gaps, and your review didn't catch them either. For the record, the scoreboard: one Gemini win, two ties, one split — a statistical dead heat. That's the last time the winner matters in this article. Here's the number that should bother you more than any leaderboard: across 700 AI-generated functions scored by the rules I'm about to use, 63% shipped a vulnerability . So "which model writes more secure code?" is mostly the wrong question — I've run that leaderboard myself and argued it's the wrong frame. But people keep asking it, so I ran it properly — on the ESLint security plugins I wrote specifically to catch these bugs, each mapped to a CWE — to show you what actually matters. The setup Four domains, four of my plugins. For each, the same feature-only prompt (no "make it secure" hint — that's how people actually use these tools), generated once by Gemini 2.5 Flash via the Gemini CLI and once by Claude Sonnet 4.6 via the Claude CLI , then linted with the domain's plugin on recommended . Method honesty: this is Gemini Flash vs Claude Sonnet — the comparable price/latency tier each vendor's CLI defaults to (Pro and Opus are a separate bracket; more on that below). It compares CLI tooling, system prompt included, not raw models under controlled decoding. n=1 per domain — but I re-ran the JWT round, and both models landed on 5 findings again with the same core misses, so treat these as directional with stable failure modes, not ±0 gospel. The scorecard Domain Prompt Plugin Gemini Claude NestJS service users + auth + admin nestjs-security 2 6 JWT auth login + verify middleware jwt 5 5 MongoDB data layer Mongoose model + search mongodb-security 8 8 General API (injection) import + search + reset secure-coding 9 13* One Gemini win, two dead h

Most agent frameworks assume one agent, one app, one bill. The moment you run agents for many clients, two problems appear that no runtime solves for you: you can't prove which client burned which tokens , and nothing stops one client's workspace from leaking into another's . I built Octorato to fix exactly that. What Octorato is Octorato is an open-source AI agent operating system: one file-native "brain" — rules, 190+ skills, 180+ specialist agents, all plain markdown under git — that a single operator runs across many sealed client "arms," with per-client token attribution and opt-in budget caps. It's not a runtime you import. It's the agent's self as files you can read, diff, fork, and own — runtime-agnostic (it runs on Claude Code today). The octopus model One brain , many arms . The brain holds the shared self: rules (the constitution), skills (HOW to do things), agents (WHO does them). Each arm is a sealed deployment serving exactly one client. Knowledge flows down (generic skills cascade to every arm) and lessons flow up (anonymized patterns get distilled back into the brain). Like a real octopus, most of the neurons live in the arms, not the head. Why "file-native" matters Your agent's identity, skills, and memory normally live trapped inside vendor code and a cloud console — you can't read the whole self, diff a change, or move it. Octorato keeps all of it as plain markdown under version control. Identity becomes diffable, reviewable, portable, and ownable . Text outlives runtimes. The part nobody else does: FinOps and isolation are the same wall Because each arm is a sealed cell that no other arm can see, every token an arm spends is attributable to exactly one client by construction. Cellular isolation is per-client FinOps — the wall that seals a client is the wall that meters it. Concretely: per-arm USD rollup (estimated from local session logs at list price), cost-spike alerts, and an opt-in PreToolUse budget gate — wire the hook and set a client's cap

I once submitted an essay with three citations that I hadn't personally verified. The AI had suggested them, and they sounded right. None of them existed. That's not a quirk or a bug — it's exactly how LLMs work. And once you understand why, a technique called RAG starts to make a lot of sense. AI assistants are remarkably good at sounding right. The model isn't lying — it's doing its best with what it knows. The problem is that what it knows has limits, and it doesn't always know where those limits are. Ask one about a recent event, a niche regulation, or anything from a source it's never seen — and it fills the gap anyway. Confidently. That's the gap RAG was built to close. Once you understand how it works, you'll have a much clearer picture of why some AI tools are genuinely reliable and others are just very convincing guessers. Here's what's actually going on. First, What's the Problem? Large language models (LLMs)—the technology powering AI assistants like ChatGPT and Claude—are trained on vast amounts of data from across the internet. That training gives them a remarkable ability to reason, summarize, and generate content. But it also comes with some real limitations: They have a knowledge cutoff. An LLM trained last year doesn't know what happened last month. They can hallucinate. When they don't know something, they don't say "I don't know"—they generate a confident-sounding answer anyway. Wrong facts, fake statistics, invented sources. All delivered with a straight face. They don't know your specific sources. Think of a software engineer asking an AI assistant about their company's internal API documentation, deployment runbooks, or architecture decisions. None of that is in the training data. The model has never seen it — and it will still try to answer. The model isn't lying — it's generating the most plausible answer it can. It just has no way to know when it's wrong. So, what do you do when you need an AI that's accurate, current, and knows your specifi

The conversation about who reads your website has been shifting. Agents are part of it now. ChatGPT fetches URLs. Perplexity reads content. Shopping agents try to complete purchases. Coding agents hit your API. Most of those products were built for humans, tested against humans. The agents showed up later and quietly. When they can't figure something out, they don't complain. They just bounce. I heard the phrase "second audience" at a hackathon where you.com was one of the hosts. It stuck. That's what agents are: a second audience the web wasn't designed for and isn't being measured against. And now, I want to build something about it. A scanner that tells you what an AI agent experiences when it tries to use your website or your API. The internal name is Perseus Clew and the public product is Agentis Lux. The split is intentional: Perseus Clew is the engine name, part of a suite of AI builder tools , and Agentis Lux is the product-facing name (Latin for "light of the agent") that describes what agent users see. This isn't a launch post. I just finished a docs phase, and I'm about to write code. Before I do, I want to put this in front of dev.to builders and find out what I'm missing. What it will do Three layers: Deterministic scanning. Twelve check categories — six for frontends, six for APIs — looking at HTML, ARIA, structured data, OpenAPI specs, error responses, idempotency patterns. Same input, same score, every time. The methodology will be published, the weights will be public, and anyone can audit it. AI-readiness scoring tools have a reputation for inflating numbers and hiding their methodology, so the trust floor is making everything inspectable. That's the foundation the rest sits on. An AI-written verdict. After the score, a Bedrock call reads the top findings and writes one sentence about what an agent experiences. Something like: "An agent visiting this page can read your product descriptions, but can't tell which button starts checkout, so it can't f

Opus 4.8 ships Dynamic Workflows — hundreds of parallel subagents per session. Read this before you wire it into prod. Anthropic's Opus 4.8 announcement on May 28 spent most of its word count on benchmarks. CursorBench up. Terminal-Bench 2.1 beats GPT-5.5. OSWorld-Verified at 82.3%. Online-Mind2Web at 84%. The legal-agent benchmark broke 10% on all-pass for the first time. Those are the numbers the headline writers grabbed. Buried under the benchmark table is the line that actually changes how you ship agents: Dynamic Workflows. Run hundreds of parallel subagents. Handle codebase-scale migrations spanning hundreds of thousands of lines. That is not a benchmark. That is a new programming model. And it is shipping as a preview, which means the defaults are not what they will be in 90 days. If you are running agents in production and you do not pin your config before the next minor release, your bill is going to surprise you. Here is what the preview actually does. Three tasks it eats alive. One class of work where it loses you money. And the exact config to pin before the dynamic-workflow defaults move under you. What Dynamic Workflows actually changed Before 4.8, parallel subagents on the Anthropic stack meant one of two things. Either you called the Agent tool from inside Claude Code and got a fixed number of side-task subagents — usually capped somewhere around four or eight concurrent. Or you wrote your own orchestrator in TypeScript or Python, called the Messages API in a Promise.all , and handled the queueing yourself. The Agent path was ergonomic but capped. The DIY path was uncapped but the orchestration was your problem — retries, structured output validation, cache invalidation, all of it. Dynamic Workflows in 4.8 collapses both. You write a script — JavaScript, not a separate orchestrator binary — that calls agent() , parallel() , pipeline() , and phase() as primitives. The runtime handles concurrency, structured output validation against JSON Schema, retri

We have watched tokens stream in from an LLM before where they appeared one at a time, like the model was typing. If you used the Anthropic SDK's .stream() method, it just worked and you probably never saw what was on the wire. This post will majorly focus on how a stream response works and how bugs are handled by SDK behind the hood. 1. Why Streaming exists To enable the streaming option we would need to make one change in the post request that is a single field "stream": true and it will change the response experience. Here are the pointers we take from the gif. The left side shows no streaming as the cursor blinks for 4 seconds then the whole response lands at once. The right side shows the streaming where the first word shows up in about 300 milliseconds. Words flow in as the model generates them. Both the sides have same model, same prompt, same total time it is just the right side started giving response almost 4 seconds earlier. The 4 seconds wait time for a full reply feels broken. A streamed reply that finishes in four seconds feels fast. Streaming doesn't make the model faster it makes the wait disappear. 2. What's on the wire When you set stream: true , the API stops sending a single JSON blob. It opens a persistent HTTP connection and pushes events down the line as the model generates them. The format is Server-Sent Events (SSE) a web standard. Any SSE debugger will read this stream. Here's what comes through: A few things to notice: The text lives in delta.text , nested inside content_block_delta events. Those are the events we should look after. stop_reason moved. In post 1 , we saw it right there in the response JSON. Here, it arrives at the very end inside a message_delta event, just before message_stop . If the loop bails out as soon as the text stops arriving we will never see it. Chunks don't line up with tokens or words. You might get "Hello" in one chunk and " world" in the next, or both in one. The network decides where the cuts happens and it

Here is what a real solo founder orders. $ git worktree list ~/app a1b2c3d [ main] ~/app-review e4f5g6h [ review-branch] ~/app-content i7j8k9l [ draft-post] Three checkouts. One machine. Each one runs its own Claude Code session that cannot touch the others. That is a normal workday for me. I run a one person shop. Content and code, same desk, same hour. Anthropic's common workflows page lists about a dozen recipes for everyday work, and the docs are strong. What they do not tell you is which recipes survive contact with a real workday and which ones stay theory. After running Claude Code as my whole operation, five workflows carry the load. Here is the honest split. https://code.claude.com/docs/en/common-workflows 1. Worktrees changed how I work The problem worktrees solve is collision. You ask Claude to fix a bug. While it edits, you want to keep building a feature. Same repo, two streams of edits, and now your working tree is a fight nobody wins. A git worktree is a second checkout of the same repo on its own branch. Claude runs inside it and never sees the other windows. claude --worktree feature-auth Real scenario from this week. The post you are reading was drafted in one worktree while a separate Claude session reviewed an open pull request in another. Neither touched the other's files. When the review finished I merged, came back to the draft, and never lost my place. If you take one workflow from the docs, take this one. The setup cost is close to nothing and parallel agents stop stepping on each other. 2. Subagents protect the one resource you cannot buy more of The model's working memory is your budget. Every file Claude reads to answer a question spends it. Ask "how does our auth refresh work" in a large repo and Claude reads a pile of files to answer. Those files now sit in the window for the rest of the session, crowding out the work you care about. Delegate that to a subagent. use a subagent to investigate how our auth system handles token refresh The