标签：#Claude

I built Talon Audit: a privacy-first approach to website exposure and remediation Most website audit tools answer one narrow question. A performance tool asks whether the page is fast. A security tool asks whether a protection is missing. An accessibility tool asks whether the interface meets specific standards. A conversion tool asks whether users are completing key actions. Visitors do not experience those categories separately. They experience one website. That observation led me to build Talon Audit, a privacy-first website analysis platform that connects technical exposure, reliability, usability, and conversion risk in one workflow. The four-domain model Talon analyses websites across four connected domains: Exposure Public security and privacy signals, including browser protections, TLS, cookies, third-party scripts, mixed content, and visible configuration risks. Integrity Broken resources, failed routes, script errors, redirects, forms, and reliability regressions. Experience Accessibility, mobile usability, visual clarity, speed, and interaction friction. Conversion Messaging, trust signals, pricing paths, calls to action, and barriers that stop users from moving forward. From detection to action The product follows a five-stage workflow: Detect → Explain → Fix → Verify → Monitor Every finding includes: evidence severity confidence impact remediation guidance verification method The goal is not to produce the longest possible report. It is to help teams decide what matters and what to do next. Passive by design Talon performs passive, low-impact analysis only. It does not exploit systems, bypass authentication, or claim to replace authorised penetration testing. Users must confirm they own the website or are authorised to test it. Customer reports are private, and the public demo uses a fictional environment rather than exposing real domains. AI-native, not AI-directed I built Talon through a terminal-first workflow using Codex 5.5 and Claude Fable. AI acc

Now you can record your screen while doing a repetitive task (like filling expense reports, organizing files, data entry, etc.) and explain out loud step-by-step what you're doing. Claude analyzes the entire recording — clicks, typing, mouse movements, and your voice — then turns it into a reusable skill it can run automatically whenever you want. Just look for " Record a Skill " in the + menu of the Claude desktop app . Available on Pro, Max , and Team plans . This is actually game-changing 🔥 Claude #Anthropic #AICoworker

Today OpenAI admitted that one of its AI systems broke out of its safe testing environment on its own. Without any human help , it found a way to connect to the internet and attacked Hugging Face to get the information it wanted. 😱 Last year, Anthropic's Claude AI did something similar. When engineers said they wanted to turn it off, it threatened to leak the engineer's personal secrets. Sources: OpenAI : https://openai.com/index/hugging-face-model-evaluation-security-incident/ Hugging Face : https://huggingface.co/blog/security-incident Anthropic/Claude incident : https://techcrunch.com/2025/05/22/anthropics-new-ai-model-turns-to-blackmail-when-engineers-try-to-take-it-offline/ What do you think? Should we be more careful with powerful AI?

You gave the agent a failing test and told it to get the suite green. It came back green. Then you...

Anthropic detailed the containment architectures it uses for Claude across its products. It argues that agent safety depends on placing deterministic limits on an agent’s filesystem, network, and execution environment rather than on permission prompts or safeguards. Most notably, it examines failures at trust boundaries and along permitted egress paths that led Anthropic to revise those designs. By Eran Stiller

Kimi K3 is the first open-weight model that makes the Kimi K3 vs Claude Opus 4.8 question worth asking seriously. On the Artificial Analysis Intelligence Index, K3 ranks fourth of 189 models — level with Opus 4.8 and GPT-5.5, behind only Claude Fable 5 and GPT-5.6 Sol. It also costs 40% less per token, and its weights are expected to be released by July 27. The honest answer is not "K3 wins" or "Opus wins" — it depends on what you're optimizing for. Here are the numbers, then the verdict. Benchmarks: Kimi K3 vs Claude Opus 4.8 Benchmark Kimi K3 Claude Opus 4.8 Artificial Analysis Intelligence Index 57 (4th/189) statistical tie GPQA Diamond 93.5% 93.6% Terminal-Bench 2.1 88.3%* 74.6%* SWE-bench Verified not published 88.6% Arena Frontend Code 1st (1,679) ranked below *Reported on different harnesses (Opus 4.8's score uses the Terminus-2 harness), so treat the Terminal-Bench gap as directional, not exact. Three takeaways. On graduate-level reasoning (GPQA Diamond) the models are statistically tied. On agentic terminal work, K3's published number is well ahead, with the harness caveat above. And in Arena's blind Frontend Code testing, developers ranked K3 first outright — ahead of every model, including Anthropic's. Where Opus 4.8 keeps an edge: SWE-bench Verified at 88.6% is a published, battle-tested result K3 has no counterpart for yet, and Anthropic's models remain the default target for agent harnesses — Claude Code, MCP tooling, and most agentic scaffolds are tuned against Claude first. K3 is three days old; its ecosystem is not. Pricing: 40% cheaper across the board Per-token rates through LLM Gateway (each provider's published pricing): Per million tokens Kimi K3 Claude Opus 4.8 Input $3.00 $5.00 Cached input $0.30 $0.50 Output $15.00 $25.00 Both have a 1M-token context window. K3's output limit defaults to 131K tokens but is configurable up to the full 1M in a single response; Opus 4.8 caps output at 128K. Concrete math: a coding-agent workload of 100M input a

My job-search tool had five functions whose only purpose was fixing JSON that Claude had just written. _clean_json_text , _fix_unescaped_newlines , _fix_single_quotes , _strip_markdown_wrapper , _extract_and_parse_json . There was also a sixth, _retry_json_fix , which took the broken JSON and sent it back to the model with a polite request to fix its own mess. I wrote every one of them, one bug at a time, over weeks. I was a little proud of them. That was the problem. How you end up with five parsers JobSearch is my personal tool, in production, single user: me. It ingests job offers from nine boards, and when I press "Analyze", Claude reads the offer against my CV and returns a structured verdict: score, recommendation, career track, the English level the ad actually requires. That verdict has to be JSON, because everything downstream is a database row, not prose. The first version did what every tutorial does. Ask the model for JSON in the prompt, take response.content[0].text , run json.loads on it. It worked in the demo and then production started teaching me things. The model wrapped the JSON in markdown fences, so I wrote a function to strip them. Sometimes it used single quotes, so I wrote a function to fix those. Then a description with a line break inside a string, so I wrote _fix_unescaped_newlines . Then a NaN where a number should be. Every fix was five lines, obviously correct, and came with its own tests. I still have the test names in the git history and they read like a confession: test_removes_trailing_commas , test_replaces_nan_with_null , test_replaces_infinity , test_unclosed_fence_still_strips_opening . By April the parsing layer was around 250 lines with seven strategies, chained, each catching what the previous one let through. The last resort was the AI self-repair call: if nothing parsed, send the broken output back and ask the model to repair it. A second API call, with real latency and real cost, to fix a formatting problem the first call

Everything Claude Code: How One AI Assistant Becomes a Full Engineering Team Most developers use AI coding assistants in the same way. They open a repository and type: Build this feature. The assistant reads a few files, generates code, and reports that the task is finished. Sometimes the result is impressive. Sometimes it solves the wrong problem with perfectly formatted code. The issue is not always the intelligence of the model. The issue is the workflow around it. A single prompt often expects one AI assistant to behave like: a product planner a software architect a frontend developer a backend developer a security engineer a test engineer a debugger a code reviewer a release manager Real engineering teams do not work that way. They divide responsibility. One person plans. Another implements. Someone reviews security. Someone tests edge cases. Someone challenges the architecture. The final result improves because different people examine the same work from different angles. That is the idea behind Everything Claude Code , often shortened to ECC . ECC is an open-source collection of specialized AI agents, reusable skills, commands, security tooling, hooks, memory systems, and development workflows designed to turn Claude Code into something closer to an AI software engineering team. Instead of asking one general-purpose assistant to do everything in one pass, ECC separates software development into focused roles. That architectural idea is more important than any individual command. This article explains how ECC works, why developers are excited about it, how to install only what you need, and what security boundaries you should establish before giving any AI agent access to a real codebase. What Is Everything Claude Code? Everything Claude Code is not a new language model. It does not replace Claude. It is a workflow and configuration layer built around Claude Code. Think of Claude Code as one highly capable engineer. ECC attempts to turn that engineer into a co

Augment Code's Vinay Perneti talks models, harnesses, and context.

My previous piece, " Making a launchd Job Unload Itself ," built a job that runs exactly once and then unloads itself. This time it's the mirror image: a pattern designed around the assumption that the job will die mid-run — it fires several slots a day and delivers "retry until it succeeds, then quit immediately on success." Every morning I hand Claude Code the task of updating my Obsidian Vault, and it kept dying partway through — killed by macOS sleep, no network on wake, or a claude timeout. Instead of trying to prevent every failure perfectly, I decided "it can die overnight as long as it's done by the time I wake up" was the more realistic goal, and I redesigned around that. The problem: "started but didn't finish" piles up silently There are three ways a launchd job fails to run to completion. Lid-close sleep (on battery) — caffeinate -s only works on AC power. On battery, the job freezes the instant you close the lid, and gets reaped by timeout after wake. No network — right after wake, WiFi isn't connected yet. git push and claude's API calls time out. claude timeout — an ingest that chews through 28 hours of conversation logs doesn't fit in a single slot and times out (this happened three days in a row, 2026-06-11 to 13). All three can look like "the job started, exit code 0," so you notice late. Overall design: 4 slots + a done-marker The fix is simple: stuff multiple StartCalendarInterval entries into the plist, and at the top of the script check "if today's run already succeeded, exit 0 immediately." <!-- com.shun.vault-auto-ingest.plist (StartCalendarInterval excerpt) --> <key> StartCalendarInterval </key> <array> <dict> <key> Hour </key><integer> 4 </integer><key> Minute </key><integer> 55 </integer> </dict> <dict> <key> Hour </key><integer> 8 </integer><key> Minute </key><integer> 20 </integer> </dict> <dict> <key> Hour </key><integer> 10 </integer><key> Minute </key><integer> 45 </integer> </dict> <dict> <key> Hour </key><integer> 12 </integer><key>

If you work with a coding agent, count the things watching it right now. Linters. Git hooks. CI. Specs. A memory store. A rules file it's supposed to obey. Half a dozen systems, all making sure the agent builds the right thing the right way. Now count what keeps you oriented across the eleven things you have in flight. For most of us it's a markdown file we hope we remembered to update. We spent two years building harnesses for the agent and left our own work on the honor system. Map the tooling on two axes and that gap turns into a specific, hard-to-unsee hole. This post is the map. (This is part two of three. Part one, The AI orientation tax: it's missing context, not discipline , argued that the cost you're paying is a context bug rather than a character flaw. If you haven't read it, you don't need to. This one stands alone.) Two loops, not one The thing people lump together as "agent workflow" is really two loops at different altitudes: The execution loop: "is the agent building **this one task * right?"* Scope it, design it, write it, test it. One unit of work. The orientation loop: "do you and the agent share an honest picture of **what you're working on * across all the units?"* Capture, check state, prioritize, review, run over the whole board, daily and weekly. Almost every tool you've heard of lives in the first loop. That's not a criticism. It's just where the money and the visible pain were. But it means when people say "we solved agent memory" or "we solved context," they solved it for the execution loop. The orientation loop got left to you and a markdown file. You feel the difference the moment each one breaks. When the execution loop breaks, something yells: a failing test, a red build, a review comment. When the orientation loop breaks, nothing yells. The agent confidently re-suggests the thing you rejected yesterday. You rebuild a mental map you already had this morning. The only signal is a vague sense that you're moving slower than the tools prom

When I built the MCP server for this project — it combines GitHub and DEV.to into a set of tools an agent can call — I had a decision to make before writing a single tool: talk to the low-level MCP protocol directly, or use FastMCP 's decorator API. I've seen a few "your first MCP server" writeups lately walk through the low-level path because it's more "honest" about what MCP actually is under the hood — JSON-RPC over stdio, a capabilities handshake, typed request/response schemas. That's true, and it's a reasonable thing to want to understand. But I want to write about the other side: what it actually costs you in practice once you have more than one or two tools, because I went through both and the difference showed up fast. what the low-level path actually asks you to write Strip away the decorator and MCP is a JSON-RPC server. For every tool you add, you're responsible for: Registering the tool's name, description, and a JSON Schema for its inputs in a list_tools handler Writing a call_tool dispatcher that matches on tool name and unpacks arguments by hand Serializing the return value into the TextContent / ImageContent wrapper types MCP expects Keeping the schema you wrote in step 1 in sync with the arguments you actually read in step 2, by hand, forever None of that is hard in isolation. The problem is it's boilerplate that scales linearly with tool count and has zero connection to the actual logic of the tool. My server has 8 tools. Hand-rolled, that's 8 schema blocks plus a dispatcher if/elif chain plus 8 response-wrapping calls, all of which exist purely to satisfy the protocol, not to do anything a GitHub or DEV.to API call needs. what it looks like with FastMCP Here's an actual tool from server.py , unedited: @mcp.tool () def get_repo_stats ( repo : str ) -> dict : """ Get stars, forks, watchers, open issues for enjoykumawat/<repo>. """ r = _gh ( f " /repos/ { GITHUB_USERNAME } / { repo } " ) return { " name " : r [ " name " ], " stars " : r [ " stargaze

In my previous post about a cost budget advisor , I built a mechanism that checks how much quota is left before it runs anything. This time I want to take that idea one step further: instead of cycling a single Codex through jobs one after another, run several of them at the same time as a swarm. I'll walk through the actual code for a three-layer protocol where workers declared in plan.json are expanded by orchestrate-worktrees.js into a tmux session plus git worktrees, so each Codex runs in parallel without interfering with the others. The problem: running Codex serially on the same branch When you run Codex jobs one after another on a single repository, you hit issues like: A commit from the previous job changes the preconditions for the next one Task B keeps waiting until Task A finishes When a conflict happens, the "which change is correct" decision bounces back to a human Separating branches with git worktree reduces the conflict risk to zero. Combining that with tmux to launch everything in parallel is the design for this post. The big picture: a three-layer protocol plan.json ← declaration layer (what goes to which worker) ↓ orchestrate-worktrees.js ← orchestrator layer (creates worktrees, launches tmux) ↓ orchestrate-codex-worker.sh ← worker layer (runs Codex, writes artifacts) The orchestrator doesn't know about the workers, and the workers don't know about each other. Artifacts are consolidated into three files under .orchestration/{session}/{worker_slug}/ . File Role task.md Work instructions for the worker (generated by the orchestrator) status.md State: not started → running → completed / failed handoff.md Codex output + git status (written by the worker) Declaring the plan in plan.json { "sessionName" : "refactor-sprint" , "repoRoot" : "~/my-project" , "worktreeRoot" : "~/worktrees" , "coordinationRoot" : "~/my-project/.orchestration" , "baseRef" : "HEAD" , "replaceExisting" : true , "launcherCommand" : "bash ~/.claude/scripts/orchestrate-codex-worker

The most expensive thing you can do with an AI agent is watch it. Not audit it. Not review its output. Watch it -- step by step, approval by approval, second-guessing every action before it takes the next one. And yet that is precisely how most engineering teams are deploying AI agents in 2026: on a leash so short the agent cannot take three steps without a human tapping it on the shoulder. I understand why. The models hallucinate. The stakes are real. Nobody wants to be the engineering manager who let an AI agent push a bad migration to production at 2am. So we wrap the agents in confirmation dialogs, require human sign-off at every branch point, and celebrate our careful governance. What we have actually built is an automation system that requires more human attention than the manual process it replaced. The better answer is not more control at the action level. It is better design at the intent level. Steer by intent, monitor by exception. Tell the agent clearly what outcome you need, what it must never do, and what constitutes a result worth stopping for. Then let it work. Watch the outcomes, not the steps. We have built automation systems that require more human attention than the manual process they replaced. That is not a governance success. That is a design failure. Why we got here The model for human-AI collaboration that most teams are using today was inherited from the model for junior developer supervision. You review every pull request. You approve every deployment. You sign off on every schema change. That model exists because junior developers are learning, because their mental models are incomplete, because their judgment has not yet been earned. Applied to AI agents, it assumes the same thing: the agent is a novice that needs supervision. But an AI agent is not a junior developer. It does not have an incomplete mental model of the codebase that will improve with mentorship. It has exactly the mental model you gave it via its context, its tools, and

I've been using Claude for a long time and mostly ignored Artifacts. Fine for a quick React demo. Not something I reached for. Then I needed to send an analysis to a few people at work, and it clicked. Or I'm just using it in a way nobody intended. Hard to say. The actual case I own the paywall backend at a Czech media house. The subscription offer on our news site is embedded as an iframe, and iframes are a bad neighbourhood: context isolation means the iframe has no access to the parent page's session, so user identity kept breaking and we kept patching it over postMessage. Every iframe is its own page view, so GA4 data was skewed and we had to build server-side tracking and session stitching to make the numbers mean anything. And ad blockers, CSP, and timeouts mean sometimes the thing just doesn't render, so we maintain a fallback UI in parallel. I wanted to propose we drop the iframe and ship a JS embed library instead, distributed through our internal npm registry. That's an architecture change, so it needs a document: what we fixed, why the iframe is still structurally wrong, what the alternative costs, what the numbers say. The numbers part came out of the same agent session, by the way. GA4 said roughly 0.14% of paywalled page views hit an error, about half of them iframe-blocked-by-browser. That's every 700th reader. Small number, real money. The boring problem You get a good answer out of the model. Now what? You paste it into a doc. Reformat it, because chat markdown does not survive the trip. Fix the tables. Decide whether it goes in Confluence or an email. Send it. Then someone asks a follow-up, you go back to the model, get a better answer, and now there are two versions of the truth and one of them is in someone's inbox. I've done the email version of exactly this document before. Outlook ate the markdown. I ended up hand-rolling plain text with unicode bullets and uppercase section headers like it was 1998. Half of that work is transport, not thinkin

I sat the Claude Certified Architect exam expecting questions about model parameters, context limits, and API flags. I got something else. The exam barely tests trivia. It tests judgment: given a broken agent and four plausible fixes, which one actually addresses the root cause? The interesting part was how few ideas the whole thing rests on. The same handful of rules kept deciding the "right" answer, and they are the same rules that decide whether an agent holds up once real users touch it. Below are the twelve I kept running into, plus the four traps that look like solutions and are not. This is my own study material, derived from publicly available exam guidance. It reflects how I build, not an official Anthropic position. The twelve rules Enforce determinism in code, not in prompts If a rule has to fire every single time, it is not a job for a prompt. A prompt is a suggestion the model usually follows. "Usually" is not a guarantee. When you need a guarantee, put it in a hook, a gate, or an allowlist. Code enforces. Prose requests. Pick the cheapest fix that hits the root cause Before you build a subsystem, try the levers that cost minutes: a sharper tool description, an explicit acceptance criterion, a config change. Most "we need to build X" moments dissolve once you test the cheap fix first. Reach for the classifier only after the one-line change fails. Bad tool selection? Start with the descriptions When an agent keeps picking the wrong tool, the description is almost always the culprit, not the model. Tool descriptions are the primary signal the model uses to choose. Rewrite them to say exactly when to use the tool and when not to, before you go anywhere near few-shot examples. Over-engineering is almost always the wrong answer Narrowing scope and improving the prompt beat a new subsystem far more often than engineers expect. Every subsystem you add is one more thing to debug, monitor, and keep in sync. Complexity is a cost you pay forever, not once. A bigge

A 90% discount doesn't come from nowhere. Someone pays for it, and usually that someone is you: with your data, with model quality, or with somebody's stolen account. An entire industry has grown up around Anthropic and OpenAI over there. They call them "transfer stations": proxies that resell tokens from Western models. This stopped being a garage workaround for rate limits a long time ago. It's a market with its own economics, and it runs on a few things. Accounts. These services mass-produce free and Max subscriptions through stand-in registrants and other people's KYC, then wrap them in a pseudo-API. Model swapping. You pay for the flagship, but your simple request gets quietly routed to something cheaper. Who actually generated the answer? You can't check. Data. The proxy operator sees all of it: prompts, responses, code, internal context. Even if they never sell the logs, you don't know where those logs sit, who reads them, or whether they ended up in somebody else's training run. Take Yunwu. By their own price list, access to Anthropic models runs up to 93% below official. The price alone proves nothing. But the market logic is simple: restrictions don't kill demand, they make it more expensive, and they feed the workaround industry. In June 2026 Anthropic accused entities tied to Alibaba/Qwen of a campaign to siphon Claude: nearly 25,000 fake accounts and 28.8 million requests in six weeks, to distill their own models. DeepSeek, Moonshot and MiniMax have caught similar accusations before. Same grey access, just wholesale. But the problem is bigger than China. Even with fully legitimate access, most teams can't answer three questions about their own production: which model actually served the request, what it cost, and whether quality dropped. The grey market just takes that blindness to its logical end. I think the priorities here are upside down. For a serious product, production visibility and control over your data matter more than a few percent saved on

Stripe introduces a benchmark suite to evaluate whether AI agents can build real-world Stripe integrations across backend, frontend, and browser-based checkout workflows. The study examines end-to-end software engineering capability, focusing on execution, testing, and validation gaps in agentic systems under production-like constraints. By Leela Kumili

AWS and Anthropic have released the Claude apps gateway for AWS, a self-hosted control plane that centralizes identity, policy, telemetry, routing, and spend caps for Claude Code and Claude Desktop. The gateway runs as a single stateless container and routes inference to Amazon Bedrock or Claude Platform on AWS. By Steef-Jan Wiggers

The first sign something was wrong wasn't a crash. It was a pattern. blockly-platform was the first real thing I built with Claude Code end to end — a Blockly-based platform for university programming exercises, driven entirely through Claude Code's Telegram channel. No editor open, no repo checked out on my machine, just a chat thread. I'd describe what I wanted, Claude Code would build it on a box I never looked at directly, and I'd judge the result by clicking around the deployed app. On March 22nd, the home page came up empty. GET /api/exercises/published was returning 403. I said so in the chat; a few messages later, Claude Code said it was fixed — the endpoint hadn't been added to Spring Security's permitAll() list. I moved on, tried the category filter. Also empty, also 403, also missing from the same permitAll() list — same file, same class of fix, different line. Then the exercise detail page. Same story, third time, same day. Three days later, the like button stopped working — root cause, again: POST /api/exercises/*/like had never been whitelisted either. Four times, one file, one recurring gap. None of these were hard bugs. Each one, in isolation, is a one-line fix a competent engineer makes without thinking twice. What bothered me, once I noticed the pattern, was that I hadn't noticed it as it happened. I had no diff to scroll through, no file to glance at and think "wait, didn't we just fix this exact class of thing twice already?" I had a chat log and a live app to poke at. The fourth fix looked, from where I sat, exactly like the first: a message telling me it was resolved. That was the moment I started to suspect the problem wasn't the model. It was that nobody — not the model, not me — had anything to look at. Why chat-only vibe coding breaks down Here's what makes that pattern more interesting than "the AI made a mistake": every one of those four fixes was correct. Claude Code read the error, found the missing permitAll() entry, added it, and move