We pick and plainly summarize new features, pricing, usage limits, and policy changes across major AI tools — Claude, ChatGPT·Codex, Gemini, and Cursor — from a solo developer and maker’s point of view.
After moving from ChatGPT to Claude, Claude’s answers can feel strong, but the old work context does not come along. Months of saved preferences, corrections, project history, and repeated instructions in ChatGPT are no longer available inside Claude. That means the same background has to be explained again from the start. Claude Projects can keep files and instructions together for one project, but that context stays inside that project. It does not automatically follow into every new chat or other AI tool, so the right project has to be chosen each time. For people who use several AI tools, the hard part is moving context once and then keeping it updated without constant manual work.
In a firsthand experience, Gemini was used to summarize forum threads, but the material seemed to be a loop of AI bots producing wrong claims and arguing with each other. If online discussions are increasingly written by bots, an AI assistant may summarize machine-made noise instead of real human knowledge. That becomes a bigger problem in an agentic web, where large language models do tasks and make choices on a person's behalf. The concern is that normal software captchas are no longer enough to tell humans and modern AI systems apart. Hardware-tied verification is suggested as a possible baseline for trusted agent requests, so AI tools do not end up acting on fake or polluted data.
Google searches for "Cursor alternatives" have risen after SpaceX acquired Cursor. The main signal is that some Cursor users may be looking at other AI coding tools after the ownership change. No exact search numbers, growth rate, time period, or named alternative tools are provided. This is not enough to prove a large user exit, but it does point to early uncertainty around Cursor.
Noam Shazeer, a co-lead of Google Gemini, is joining OpenAI. Gemini is Google’s main family of generative AI products, while OpenAI makes ChatGPT and Codex. The move shows that Google and OpenAI are competing directly for the people who shape major AI models and developer tools. No specific product changes for ChatGPT, Codex, or Gemini have been announced yet.
Papercrane CLI is a command-line tool that lets Claude Code connect to business data, build a dashboard from a question, and publish it as a shareable link. Dashboards made by AI coding tools often become static files with numbers fixed at the moment they were created, but Papercrane puts the data connection inside the dashboard so the figures can keep updating after the chat is over. It connects to more than 50 services, including GA4, Stripe, HubSpot, Salesforce, Postgres, and BigQuery, through normal browser sign-in instead of pasted secret keys. Credentials are encrypted by Papercrane, shared at the organization level, logged for review, and removable from one central place. `papercrane publish` turns the dashboard into a hosted link with access control, embedding, and custom domain options. A dashboard can also write changes back to connected tools, such as changing a Salesforce deal stage from inside the dashboard. If a needed connection is missing, the agent can create one as a TypeScript handler, and the dashboard code remains in the user’s workspace as Next.js and React source that can be pushed to GitHub. The tool is built around Claude Code, but Papercrane says Codex, Cursor, and other AI coding tools that can read documentation and command output can use it too.
A parent with children aged 7 and 12 had a constant fight over Roblox, especially Rivals. Blocking it, limiting it, and setting timers only turned play time into a daily struggle. The approach changed from stopping the game to building a first-person shooter together. The kids chose the features like product managers, while the parent and Claude handled the coding. After naming the game Cooked, they had a playable browser game in the first 3 hours. The next day, two more children joined, and the group used 3 laptops, 1 iPad, and 1 iPhone. Mobile support was added in about 30 minutes so all 5 people could play together. The kids asked for features like a rocket launcher that pushes players off the map and a knife that attacks faster when held, then those ideas were turned into prompts that Claude implemented.
Long Claude Code sessions can become unreliable when the context window fills up. A person can see the remaining context through the on-screen display or by running `/context`, but Claude Code cannot run that built-in command by itself. `/claude-context` is different because it is a skill, so Claude Code can call it during work and read its own context status. That makes it possible to build safer workflows, such as compacting earlier when the session gets crowded or taking action before five-hour and seven-day limits matter. Claude Code still cannot start a fresh session on its own, so a more practical pattern is to delegate work to subagents while the parent agent watches the overall context. The wider pattern is the same across recent tools and habits: coding agents need cleaner memory, shorter project notes, decision logs, less terminal noise, and better access to web pages or repo knowledge so they do not lose the reasons behind earlier choices.
SubChat is a browser extension that lets people ask a small follow-up question beside a specific part of a Claude answer. The problem is that a long Claude chat can become hard to follow when a new question pushes the whole conversation forward. With SubChat, a person can highlight a line, press an ask button, and get a separate answer in a movable side window. Several side chats can stay open at the same time, and answers or code can be copied quickly. It uses the existing Claude login session, so no separate API key is needed. It also says it uses no servers, analytics, or tracking, and it is shared with a GitHub repository.
Linus Torvalds strongly rejects the idea that AI should be described as writing nearly all of someone’s code. In his view, AI coding tools are productivity tools, much like a compiler. A compiler made programming far faster, but people do not say the compiler wrote their code, so AI should not be described in a way that removes human responsibility and judgment. He is not against AI and appears to use and understand it. The Linux kernel reportedly saw a 20% rise in submissions during this release cycle because of AI tools. The downside is hitting small open source projects. AI-generated bug reports are increasing, and some people disappear when maintainers ask for more details or a patch, leaving one- or two-person projects with extra work and more burnout risk.
In one firsthand business setting outside a major city, Claude is used every day for work. It helps with operations, planning, and business development, and it is also used on weekends to vibe-code apps and websites. The tool has helped the business grow, with more learning expected ahead. Many business-minded people nearby still do not know how AI tools could help them. The rough personal estimate is that about 1 in 20 people use AI lightly for productivity, while about 1 in 100 are learning to use it as a real business tool. The gap changes the feeling of being behind: compared with many people offline, regular Claude users may already be ahead.
OpenAI’s Codex CLI has a reported logging problem that can create very large SSD writes when it is left running for long periods. The tool keeps writing diagnostic logs into a local SQLite file at `~/.codex/logs_2.sqlite`. A June 14 GitHub issue measured about 37 TB of writes over 21 days, which works out to roughly 640 TB in a year. A typical 1 TB consumer SSD is often rated for about 600 TBW, so this level of writing could use up the drive’s warranted write life in under a year. The likely cause is a logging setup that runs at the very noisy TRACE level by default and records things like WebSocket data and ordinary file access events. The usual RUST_LOG setting does not appear to quiet it easily, and about 71% of the logged data was described as low-value TRACE noise. The problem is worse than the visible file size suggests because the database is doing many insert and delete operations per minute, causing write amplification. A temporary Linux and macOS workaround is to symlink `~/.codex/logs_2.sqlite` to `/tmp/`, which redirects the writes to RAM; the file reportedly does not store conversation data, so losing it after reboot is acceptable.
EvenKeel is a free chatbot for budgeting, financial planning, retirement saving, and tax planning. It runs on Gemini 3 Flash Preview and includes ready-made abilities for tasks such as checking whether a home purchase makes sense, planning insurance, and choosing a mix of investments. It can also do financial math and Monte Carlo simulations. Users can upload files or explain their situation in chat. As the conversation continues, EvenKeel keeps updating a financial picture with key facts such as the user’s situation and goals. That financial picture, along with recent chat messages, is sent back to the model each turn so important details stay available. Users can export their data or delete their account from settings, and they can also use it anonymously until model costs become too high.
Anthropic released a Swift package that lets Apple apps use Claude through Apple’s Foundation Models framework. Claude can act like a server-based language model inside the same system Apple uses for its on-device model. Developers can use the same LanguageModelSession flow to ask Claude for answers, stream replies, request structured output, and use tool calling. Requests go from the app to the Claude API, and Apple does not see the prompts or responses. Billing goes to the developer’s Anthropic account under normal API pricing. During development, an app can use a Claude API key directly, but shipped apps should use a proxy because a key inside an app can be extracted and abused. Apple’s on-device model is fast, private, and works offline, while Claude is meant for harder tasks that need longer context, stronger reasoning, web search, or code execution. The package is still beta and targets iOS 27, macOS 27, visionOS 27, watchOS 27, and Xcode 27 betas, so the API may change before full release.
In one firsthand case, Claude-like AI tools have sharply increased the speed of solo software work. An automated site was built to find and publish only positive news about the UK while running on a cheap VPS. Claude wrote the site, and free tiers on several AI platforms checked news feeds, picked useful items, and rewrote clickbait-style headlines in a calmer form. The site links back to the original articles instead of storing full copies, which keeps readers going to the publishers. A second project rebuilt an internal company video hosting tool. That tool stores playable master videos, edits, and exports for editors, production staff, and broadcasters, like a low-cost Vimeo. The old system was no longer good enough for 2026, and the goal was to add features similar to Vimeo, Frame.io for review, and Trint for transcription. Fable and Opus rewrote a decade-old PHP, Ajax, and nginx setup into a Python system using Django, React, and Vite.
FactIQ is an economy and finance data plugin for Claude Code and Codex. It helps an AI agent spend less of its limited working space on finding and cleaning data, and more on analysis. It brings together SEC filings, official statistics from the United States, China, India, Korea, IMF and World Bank data, live market data, and earnings-call information. A free FactIQ account is enough; no separate codebase or self-hosted database is required. The agent can find the right data series, run read-only SQL, calculate new measures, and produce a shareable chart, report, terminal preview, or local HTML visualization. The data is reshaped from about 20 official sources into three main tables: `series`, `data_points`, and `dimensions`. Claude Code uses a marketplace plugin install, while Codex uses `codex plugin` commands and an OAuth login to the MCP server. Economy or finance questions can call the skill automatically once it is installed.
A licensed doctor who finished residency is using Claude to prepare for board exams. Claude has helped with medical learning, early work as an attending, and everyday tasks, so it has become a tool that understands their thinking style. When they ask Claude to explain practice questions from board review banks, the system often marks the question as sensitive and moves the request away from the chosen model. The request may be sent between Sonnet, Opus, and Haiku without a clear pattern, and the answer quality changes enough to notice. Sometimes similar questions stay on the newest chosen model, which makes the routing feel inconsistent. The doctor would be willing to prove their credentials if that allowed better handling of medical education questions. The main concern is that strict safeguards can block legitimate learning by healthcare workers, which may also affect the patients they later treat. Because the doctor has ADHD, Claude’s ability to keep their thinking focused on the medical concept was more useful than a normal written explanation.
A solo founder added short voice guidance to a breathing meditation feature in a wellness app. The feature already had a breathing animation, calm background music, and a timer. The missing part was simple spoken guidance such as inhale, hold, and exhale. Because OpenAI was already connected to the app, the founder used OpenAI’s text-to-speech API to create audio files for the welcome, inhale, hold, exhale, and finish prompts. Those files were then saved inside the app and played directly during the meditation. The app does not contact OpenAI during each session, so the feature works without internet, starts faster, and does not create ongoing usage costs. The main idea is using AI once during the build process, instead of running AI live inside the product.
PATCHR-Studio is a free, open source app built with Claude Code to make protein structure work easier for lab researchers. AI for protein structure prediction has become powerful, but many advanced tools still require command-line setup, memorized options, and little visual feedback. Easier commercial tools can cost thousands of dollars per year, with the cheapest academic license mentioned starting at $7,500 per year. PATCHR-Studio shows a protein structure in 3D and lets people rotate, zoom, click, and edit it with a mouse. It can fill missing parts, change a chosen spot and recalculate nearby changes, remove a part, add a modification, and export the result for a simulation. It runs on macOS, Windows, and Linux, and the GitHub repo is DeepFoldProtein/patchr. By the builder’s estimate, it took 1 to 2 months, mostly spent on user experience iteration, with no separate API bill beyond a Claude Max subscription. It covers about half of the expensive flagship product’s day-to-day features; of the remaining gap, some work is seen as company-scale engineering and the rest as details that need more work.
Composer 2.5 Fast in Cursor works well enough for coding and browser testing, but it is less satisfying when writing instructions or answering questions clearly. The answers can feel too wordy and not clear enough. Its explanations of what it did are also not strong enough for that use. Because of this, it is being used more like a subagent for specific tasks, rather than as the main assistant. API credits are nearly used up, so better setup tips or a testing harness would be useful before spending more.
The author grew up writing toy kernels and OS projects as a way of learning C, including a project called MontaukOS started in early 2025, where much of the core kernel code was written by hand. The project sat dormant for months after other priorities took over. When Claude Opus 4.6 was released, the author was introduced to Claude Code and agentic AI coding for the first time, and decided to pick the dormant project back up to test how well the model handled low-level code. The results were far better than expected, and over time the author ended up writing very little code by hand. The project grew from a toy kernel that did very little into something dual-booted and used daily on a laptop. Claude Opus 4.6, among other models, produced a full networking stack, disk and filesystem drivers, an Intel graphics driver, a desktop environment (something the author says they could never have built due to weak graphics-coding skills), and a PDF viewer.
Meta made a large workforce shift earlier this year to speed up AI development. The company laid off about 8,000 workers and moved about 7,000 people into AI-focused teams. The change was meant to support Meta’s AI infrastructure spending, which could reach $145 billion in 2026. Mark Zuckerberg told employees that Meta’s AI agents have not improved as quickly as expected, and that leaders misjudged the timing and rollout of the changes. Meta had been highly optimistic about tools like Claude Code earlier this year, but agent progress over the past few months did not speed up the way executives expected. Zuckerberg still expects Meta to see clearer benefits from its AI spending within three to six months. Meta also paused a controversial program that tracked employee mouse and keyboard activity, and plans to make it opt-in if it returns. CTO Andrew Bosworth acknowledged that the launch of the Applied AI group was handled badly and damaged trust because teams were moved quickly without a clear enough explanation.
Cursor saves chat transcripts on your computer, but they are usually separated by project and are hard to search together. memgrep collects Cursor agent transcripts into one local searchable memory. It can also work with Claude Code and Kiro histories. `npx memgrep ingest` scans the chat history, `npx memgrep recall "how did we fix the auth race?"` searches past chats by meaning, and `npx memgrep copy` copies the best result to the clipboard so it can be pasted into another chat. The first recall or ingest downloads an embedding model of about 25MB, and after that the tool runs offline. As an MCP server, memgrep can let a Cursor agent look up old chats while it is working, including chats from other projects. The agent gets `recall(query)`, `get_chat(id)`, and `list_chats()` tools for finding and opening past conversations.
The main issue is how much setup to do before starting a new project in Cursor. Common advice says to write Cursor rules, add docs, describe the app structure, define the code style, and keep the agent's context focused. That advice can help, but beginners can get stuck preparing the agent before there is any real app shape to work from. The practical question is what context is worth setting up before the first build, and what should wait until the project starts showing real problems.
A developer on vacation had to send a client an invoice immediately. The client needed it for end-of-fiscal-year accounting, and waiting until the evening was not acceptable. The developer had no laptop and only had a phone while standing in a roller coaster line with family. There were about 20 minutes before the ride. The developer opened the Claude app, attached an older invoice PDF, and asked Claude to recreate it as a new invoice. Claude wrote and ran a Python script that made a close copy of the original PDF. After being told to change the invoice number, billing period, and costs, Claude generated the needed PDF. The developer attached it to an email and sent it just before getting on the ride.
In AI work with many documents, a strong system prompt still fails if the model is looking at the wrong part of the file. Rewriting instructions like “answer only from the provided text” does not fix the deeper problem when the provided text is a large, messy paste from a PDF, spec, or markdown note. The better setup separates context retrieval from prompting. Linkly AI was placed over MCP so the AI agent can see folders as smaller, searchable pieces instead of one huge block of text. The agent checks the folder map, looks at a document outline, and reads only the exact section it needs. Local files stay local, and the prompt becomes shorter and stricter. The main instruction becomes: use the retrieved section as the source of truth, cite it, and do not hallucinate beyond it. The retrieval layer now works well, but the remaining open issue is how to design the system prompt around it.
claude-real-video, also called crv, is an open-source tool that prepares videos for Claude. Claude cannot take a video file directly, and a YouTube link usually gives it only the transcript. crv finds the frames where the scene actually changes, instead of grabbing one image every second. It also turns the audio into text and saves everything in a folder Claude can read. All processing runs locally on the user’s computer. Version 0.7.0 adds crv-web, a local web page in English and Chinese where a link can be pasted, analyzed, and viewed without using the terminal. It can be installed with pip install claude-real-video and is released under the MIT license. The GitHub project has passed 1,300 stars and reached the Hacker News front page a few weeks earlier.
DwarfStar, or ds4, is a local AI inference engine released on GitHub on May 7, 2026 by Salvatore Sanfilippo, the creator of Redis. It is written in C and built for one specific model: DeepSeek V4 Flash. That model has 284 billion total parameters, with 13 billion active during use, and uses a Mixture-of-Experts design. On machines with a lot of memory, ds4 also supports a PRO version with 1.6 trillion total parameters and about 49 billion active parameters. ds4 is not a general Ollama replacement or a wrapper for GGUF files. It is a purpose-built engine meant to run one frontier-level model as efficiently as possible on consumer hardware. Its main target is Apple Metal, with support for NVIDIA CUDA and AMD ROCm. The normal memory target is 96GB, but SSD streaming can lower that target to 64GB.
After several months of using Cursor almost every day to build client websites, the hard part is not getting usable code. The work includes small business sites, single-page sales pages, redesigns, and portfolio sites. The setup already includes project rules, reusable site parts, written notes, favorite examples, and stored company knowledge, so each job does not start from nothing. The current design workflow is to capture screenshots of a liked site or app, ask Cursor to keep the same visual style, and adapt it to the new project. Cursor can sometimes get close, but often misses the small visual details that make a page feel polished: spacing, type choices, visual priority, buttons, proportions, and overall feel. The result becomes a long back-and-forth with prompts about adding more empty space, fixing the first screen, making buttons less generic, reducing noise, and moving closer to the reference. The final result can be acceptable, but it takes more rounds than expected. Cursor also seems weak at judging its own screen design.
Frequent Claude Code use can make coding feel easier while also creating fear that core skills are weakening. The concern is that reading code, writing code, and thinking through software problems may get worse with daily reliance on the tool. A more cautious workflow is to first decide the broad solution, then plan the code design, and only use Claude Code when needed. The tradeoff is that output can feel slower compared with developers who lean heavily on AI assistants. The deeper question is whether software engineers should remain people who solve and understand problems, not only code reviewers for AI-written code.
Long AI chats can become too large, slow, and confusing to keep using. The work is not lost, but useful decisions, notes, and project details can get buried inside a messy conversation. This desktop tool takes copied chat text and turns it into a cleaner project structure that a new AI chat can continue from. It can detect separate projects or topics, let the person review and fix the structure, remove noise, and merge or split projects. It can export a full rescue package, a single project handoff, a reference file, or a ZIP archive. The goal is not a normal summary of what happened, but a usable working state so the person does not have to start again from zero.