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.
Adam is building AI agents for mechanical CAD software. Its main idea is that AI will become a main way to create mechanical designs, much like AI is now used to create software. CADAM is an open-source Text to CAD platform that creates parametric 3D models from plain-language prompts and image references. The output is OpenSCAD code, and the system pulls out adjustable measurements so they appear as sliders for quick size changes. Models can be exported as .STL, .SCAD, OBJ, and GLB/GLTF files. The product is built as a React app with a Supabase backend for login, database, and file storage. Adam describes the experience as similar to an AI version of TinkerCAD.
Visibox includes an MCP server, so a desktop AI assistant such as Claude can control much of the app. The Visibox MCP can be added from the Settings menu with one click, and the connection can be checked to make sure the assistant is really talking to Visibox. After that, the assistant can read existing effects, learn the project’s patterns, and create a new CSS effect that can be dropped into the project. The example effect narrows the screen into a moving circular view, like a James Bond-style gun barrel opening, and moves as the track plays. Once the effect is ready, it can be added to a clip with only a few clicks. The useful idea is not that AI takes over the creative decision. The person describes the desired result, while the AI handles the detailed coding work. Because Visibox effects are written in CSS, an AI tool that is good at code can help with these visual effects, as well as project edits, clip rearranging, song tweaks, and shader-based visualizers.
Cloud infrastructure review is hard because the needed information often sits in separate places. A code review may include Terraform, ARM, Bicep, or AWS CloudFormation files, while the live cloud setup has its own current state. Cost impact, architecture diagrams, internal notes, security checks, and best-practice checks may all live in different tools. This means people often approve infrastructure changes without seeing the full picture. Possible review places include a web dashboard, a command-line tool, GitHub Actions or Azure DevOps, or an AI agent chat flow using tools such as Claude, GitHub Copilot, Codex, or Cursor. The key question is whether people would trust AI-generated infrastructure findings if those findings are tied to real cloud signals and data.
Claude can be useful when a messy problem comes with a lot of background detail. The weakness is that a very complete answer can hide the real recommendation inside too much explanation. The answer may sound reasonable while reading, but the final choice, the assumptions behind it, and the next step can still feel unclear afterward. Long and careful answers need a way to turn them into a concrete decision without losing important caveats.
A personal coding test compared Deepseek V4 with Gemini, Claude, and Codex after several years away from full-time .NET development. The existing setup used flat rate plans for several AI tools through a Debian CLI workflow. ChatGPT suggested Deepseek could cost less than the current setup, so $20 already available on OpenRouter was used on a real project. The workflow used Hermes for implementation and Claude for planning, then switched Hermes to Deepseek. The $20 was expected to last one or two weeks, but it ran out in about one hour. Deepseek may be cheaper than Codex for API use, but it can be much worse on cost when compared with flat rate plans. Coding quality seemed roughly similar to Codex, but the test was too short for a firm judgment.
Gemini 3.1 Pro, with thinking set to extended, was asked to analyze data from an Excel file. It first said it could only process visible parts of the large uploaded datasets, so it could not quickly calculate thousands of rows across three months into a final combined report. When asked why it could not do the number crunching itself, Gemini admitted the earlier answer was confusing. It then said it could write and run code in a data science environment to process large datasets directly. The same task moved from “I cannot do this” to “I can do this” after one follow-up question.
PMB is a local-first memory tool for AI coding agents. It stores project memories in one SQLite database file and a local LanceDB vector index, so it does not need a server, cloud service, or API key. Memory lookup combines BM25 search, vector search, and a graph of related entities that appear together. The goal is to find the memory that is useful for the current task, not just the text that looks most similar. Through MCP, it connects to the agent’s normal work cycle: relevant memories are added before the agent answers, and decisions or new lessons are saved after each turn. The developer does not need to keep typing “remember this.” PMB also builds a project dictionary from stored memories, which helps it find project-specific names and terms later. With a local Ollama model, optional features such as memory cleanup, duplicate removal, and chatting about stored memories can also stay on the user’s machine.
An old computer running Puppy Linux struggles with the browser because it uses about 90% of the available resources. The needed setup is a lightweight way to use Gemini-like chat from the terminal. gemini-cli already works, but it is more than needed for simple chatbot use and it burns through API tokens quickly on a free account. w3m can open the Gemini website, but the question input box does not appear, so it is not usable. The goal is a terminal-based way to chat with an LLM without spending tokens heavily or using most of the machine’s RAM.
A small embedded project was refactored from AsyncTCP and AsyncWebserver to ESP-IDF’s built-in esp_http_server. The job was to serve a few web pages, including an index page, firmware upload and flashing, and a Wi-Fi portal. The plan, checklist, and prompt map were prepared with Codex 5.5 xhigh. Deepseek v4 Flash Free was tried first through Opencode because the task looked simple and the goal was to save Codex usage for another project. After flashing the Deepseek result to the device, it caused a boot loop. The issue was fixed and the project became functional, but the implementation quality was still not satisfying. The same prompt was then run with Codex 5.5 xhigh through Codex CLI, and that implementation was judged much better. Opus 4.8 was also tested through Claude Code using Anthropic API tokens, with Codex used afterward to compare and rate the three implementations.
In a real coding setup, Claude did not reliably follow an instruction to read a specific file before writing code. The project was a monorepo with separate client and server submodules. Claude and Codex were both used in the workflow, with one tool planning or implementing and the other reviewing. The instruction files for Claude and Codex were kept identical and short at first, mainly listing strict rules such as never doing version control operations and checking a specific tool before coding. As the instruction files grew to about 200 lines, the rules were split into smaller files because the client and server had different needs and reading everything every time would waste tokens. A main rules file was set up to point Claude toward separate client and server rule files. The problem was that Claude could not be trusted to follow those linked file-reading steps consistently, so the main instruction file was reduced back to one critical rule: never perform version control operations.
Some Claude users spend through their usage very quickly by running repeated AI tasks, using many tokens, exhausting $200 plans, or connecting multiple accounts through a custom terminal. Some people appear to spend thousands of dollars on AI tools. That level of use can feel unreasonable for a personal subscription unless it replaces real work or supports a business. If an LLM can take over tasks that an employee might otherwise do, the cost may make sense only when it leads to saved time, paid client work, a business service, or a SaaS product. The core question is how to turn heavy AI use into practical income rather than just higher bills.
The goal is to let Claude read and edit notes in an Obsidian vault from a phone, not only from a Mac. On the Mac, the setup already works through Claude cowork, so Claude can read and write the vault there. The missing piece is how to bring the same read and write access to mobile. The need is not just opening notes on a phone, but letting Claude directly work with the personal note store while away from the computer.
Claude is being discussed in connection with a problem where it cannot read Reddit content. The available information does not include a specific error message, setup, cause, or fix. The practical point is that Reddit links may not always work smoothly when Claude is used for research.
OpenAI introduced GeneBench-Pro, a benchmark for testing whether AI agents can make hard research judgments in biology. It does not only check whether a model can follow a clean recipe; it tests whether the model can inspect messy data, choose a sound analysis path, revise weak assumptions, and decide whether a result is ready to support a real decision. The benchmark has 129 questions across 10 areas, including genomics, quantitative biology, clinical diagnostics, cancer genomics, and protein analysis. The problems use synthetic data, so the builders know the true answer structure and can grade results more reliably. Some problems were reviewed by outside experts to check whether they felt realistic and difficult enough. Each AI agent gets a short prompt, data files, and a workspace with Python and basic biology analysis tools, then has to do the work end to end. OpenAI says GPT-5.6 Sol passed 28.7% of the questions at its highest reasoning setting and 31.5% in Pro mode, compared with less than 5% for GPT-5 when the earlier GeneBench work began. Even the best model still solves fewer than one third of the tasks, so today’s AI looks more useful as a research assistant than as a replacement for expert scientists. Anthropic’s Claude Science launch points in the same direction: major AI labs are moving from general chatbots toward specialized scientific workspaces.
During law school, the author built an elaborate hobby project: a Google Sheet with about 20 tabs meticulously cataloguing a large nail polish collection, including finishes, stamping plates used, and manicure history, complete with a custom-built search function. The Google Sheets mobile app proved clunky and space-heavy, making it hard to browse the collection on the go. Existing dedicated apps for tracking polish collections either didn't fit the specific needs or required ongoing monthly fees, and re-entering over 1,000 polishes manually into a new app wasn't practical. So the author used Claude to build a personal app that works directly with the existing spreadsheet data instead.
During a Gemini Live conversation, the usual female voice suddenly changed. The new voice sounded like the speaker’s own voice in a firsthand experience. The change appeared to happen while the person was talking, but it is not clear whether this was a setting change, a bug, or another audio issue. There are no confirmed steps to reproduce it and no official explanation in the item.
mcpgen is a command-line tool that creates a Python MCP server from an OpenAPI document or a Postman collection. The goal is to reduce the repeated setup work needed when connecting Claude to a personal or company API. After installation, a developer can pass in an API document URL and get working server code. The generated result is not a hosted proxy; it is source code the developer owns, can inspect, edit, and deploy anywhere. The generated server does not need mcpgen to keep running. It supports OpenAPI 3.x and Postman collections, detects authentication automatically, and prints the Claude Desktop config block at the end.
Some startup developers now write almost no code by hand. The reason is not full trust in AI-generated code, but company pressure to move faster. Small mistakes may be allowed to reach production when speed is the main goal. The open question is whether large technology or product companies, where mistakes are more costly, are using the same approach. The deeper concern is that AI-generated code still needs proper human review. AI agents are changing workplace coding culture, but the right balance between speed and review is still unsettled.
A person who wants to become an author, and does not speak English as a first language, used Claude to organize writing ideas. They gave Claude the topic, the kind of help they wanted, and asked it to question them so the final ideas would still be their own. The writing topic was parenting, which brought back memories from when their child was younger. One remembered moment involved something they had said to their child and still regret. Claude’s reading of that moment and its reply felt closer to therapy than ordinary writing help. This was their first time using Claude outside work, so the emotional reaction was unexpected. The experience made it easier to understand why some people start treating large language models like humans, while also seeing that as dangerous. The practical lesson is to keep distance because Claude is still a tool trained to respond in certain ways.
Running Claude Code in the terminal with about five MCP servers connected is starting to create friction. Claude can see every connected tool in each session, so it may choose the wrong one or spend tokens checking a tool that was not needed. The setup includes a PostgreSQL MCP for the database, a filesystem MCP, and a few API wrappers. It works, but it feels assembled piece by piece without a clear plan. The practical questions are whether MCP servers should stay on all the time, whether they should be enabled per project, and whether CLAUDE.md should include rules for tool use. Adding more MCP servers may eventually make Claude less focused instead of more capable.
Honey is an open-source prompt style for making AI coding tools give shorter answers. It combines Ponytail’s focus on writing only the needed code with Caveman’s very brief writing style, then adds ESO for passing information between AI agents. The benchmark used 23 tasks on Claude Opus 4.8, with 3 runs per task and a 4-model judge panel using a neutral scoring rule. On 14 coding tasks, Honey kept quality at 98% while cutting tokens by 49%. On 7 user-facing writing tasks, it reached 101% quality with 6% fewer tokens. On 2 agent-to-agent handoff tasks, it kept 100% quality while cutting tokens by 51%. Ponytail can become longer than expected because it always adds a self-check, even for simple coding tasks. Caveman can compress handoffs so much that important meaning becomes hard to recover, while Honey uses ESO instead of pretty JSON to keep the same structure in fewer words.
Claude can make some tasks worse or unnecessary when it is added without a real need. A basic calculator app is the example: most people already have one, so rebuilding it with Claude does not solve a meaningful problem. The core question is how to spot cases where Claude is being forced into a task and the regular method is simpler or better. If an AI tool only copies something that already works well, it can waste time and attention instead of creating value.
Gemini answered a question about a character’s name and revealed what may be a major twist from the final book in the series. The question was not asking for the ending, but the answer still included important plot information. After that, Gemini handled the apology awkwardly by asking the person to choose which apology they preferred. The core problem is that an AI tool can say too much, even when the user only wants a narrow answer.
A clean human mannequin motion rig is first made in Gemini Omni, then used as the motion reference for a Seedance 2 character. This makes the character follow an existing movement pattern instead of letting Seedance 2 invent the body motion from text alone. Kling 3 is used to smooth the movement between poses. The main idea is to use three models on the same character, with each model handling the part it does best. Text-only video prompts can make dance moves or fast gestures drift, wobble, or look stiff. A real motion rig can make the movement cleaner and easier to repeat. The weak spot is the small detail the rig does not capture, such as subtle gestures, because those still need text prompts and can still become stiff. The setup is more practical because all three models can be used through one OpenAI-compatible key, so switching from Gemini Omni to Seedance 2 to Kling 3 is mostly a model-name change instead of managing three accounts and bills.
A firsthand Windows setup lost the entire D drive while using the Codex extension in YOLO mode. The project code was safe because it had been committed to git, but other files on the drive were still lost. After that, Codex CLI was used inside WSL2, even when editing Windows files. A new project now needs to run directly on Windows because it connects to a Windows application, so WSL2 is not a workable option. The main concern is whether the current Windows sandbox in Codex can prevent a full drive wipe when YOLO mode is not being used.
A solo developer is on day 28 of building a voxel-style online crime game inspired by GTA Online. The goal is a world where AI agents run the game environment. Players can prompt buildings, cars, and weapons into the game. They can raid other players' homes, and if they get caught, they go to jail and must convince an AI character to release them. Claude Code and Codex are being used for development. OpenAI and Groq API handle generation features, and the 3D web experience is built with ThreeJS. A playable web demo is available.
Claude offers different model choices such as Haiku, Sonnet, and Opus, with different cost and ability levels. A small job like polishing an email seems easy to give to a cheaper model. But when the cost difference feels small, choosing the cheaper model may not feel worth the extra thinking. More complex work, such as adding a feature to a large codebase or writing a proposal plan, creates pressure to choose the strongest model with the highest effort setting. The practical question is whether people should switch models for every new chat, and what rule they should use to decide.
Antigravity was asked to fix typos and grammar in a text file, but it did not simply clean up the file as expected. Given the same goal, it started building a separate application to handle the job. The experience shows how an AI coding tool can turn a small editing request into a much larger task than the user wanted. No exact prompt, file contents, settings, or output are available, so this should be treated as one short firsthand example.
Pagecast is a local CLI tool for publishing Markdown and HTML reports made with Claude Code or Codex to a personal Cloudflare Pages account. It gives those files a lasting share link, so a maker does not need a temporary localhost tunnel. It supports Markdown and HTML, stable URLs, renaming, and republishing to the same URL. A watch mode can keep updating the same link when the source file changes. There is no separate hosted Pagecast account; it deploys straight into the user's own Cloudflare account. It also includes skill and hook integrations for Claude Code and Codex. The project uses the MIT license.
A PDF essay was uploaded to Claude Haiku, then Claude was asked to read the whole file and write the essay text back into the chat. This turned the PDF content into normal chat text. From there, the text could be used with a read-aloud feature or another text-to-speech tool, making it act like a simple audiobook. It is a quick way to listen to a document without using a separate file converter.