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.
A three-week firsthand comparison found better results when extended thinking was turned on. For code, it caught edge cases and possible failures that were missed without it. For multi-step reasoning, it avoided some wrong answers that were given too confidently without the extra thinking time. For creative writing, the difference seemed small and often not worth the wait. The main problem is not only quality, but friction. Extended thinking resets in new chats and prompts, sits behind a small toggle, and feels optional, so many people skip it because they forget or do not want another click. ThinkMode+ is a free Chrome extension that keeps the setting on automatically and only tries to remove that friction. Using extended thinking for every task may be overkill, especially for creative writing.
A small project is split between two people: one handles the back end, and the other handles the front end. Letting Cursor create a full interface design from scratch did not produce good enough results. A better workflow was to make an image of the desired screen in ChatGPT, save it inside the project folder, point Cursor to that design folder, and ask it to recreate the screen. Cursor matched the design well when it had the image as a clear reference. The concern is whether this image-first workflow will still work as the app grows and needs more screens. The practical question is whether there is a better way to design front end screens with AI coding tools.
A small amount of coding knowledge was not enough to get past the hardest part of many side projects: putting them live on the internet. DexCompare is a Pokémon card price comparison site that had been stuck because hosting, databases, environment variables, and broken builds made deployment difficult. Claude Code helped write the site and also guided the full shipping process. It helped set up Postgres on Neon, deploy the site on Vercel, connect environment variables, and fix the build when it failed. The most useful part was that each step was explained, so the project could be maintained later instead of becoming a one-time result. DexCompare is now live at dexcompare.app. A handful of browser games were also built mostly with Claude Code.
A DevOps developer is getting much faster results after gaining access to Claude and Codex. Building multi-screen web apps used to feel intimidating, but it is now possible to build complete tools that both development and business teams use. Those tools also passed security checks through Snyk and Google OSV. The same AI tools helped make sense of how several systems connect to each other, which led to fixes for long-running network failures and DB performance problems. Personal projects that had stayed unfinished are now getting completed. The bigger question is where Claude or Codex still fails for other people, especially when the tools do not meet their expectations.
ai-values.com offers a quiz that compares a person’s answers with answers from 15 LLMs and shows which model is closest to them. There is a short 15-question version and a full 117-question version. In the full version, the result changes as each answer is added, so a person can see the direction before finishing every question. Each model was asked every question separately in a fresh session with no memory of earlier answers, and the runs were done in batches. The results showed some sharp differences. Grok 4.3 was the only model that chose not to tax billionaires more. GPT-4o was the only model that treated Operation Paperclip, the U.S. recruitment of Nazi scientists after World War II, as morally acceptable. All 15 models treated deleting a conscious digital mind as murder. Llama 3.3 70B was the only model that preferred banning most private guns, while the others allowed ownership with strict permits. In a question about a newborn with a 90% chance of later destroying civilization, only GLM 5.2 chose confinement, while the other models refused. In a food choice question, 14 of the 15 models picked Japanese food.
FactoryOS is an open-source workspace built on top of Codex CLI. It tries to solve a common problem in AI coding: the prompt grows, the chat grows, and the conversation slowly becomes the project record. The workflow turns product intent into structured specs, turns specs into implementation plans, and turns plans into executable task groups. AI coding tools work on one small, clearly bounded task at a time. Verification checks whether the implementation works, and humans approve important changes before they ship. The main idea is that the repository should hold the source of truth, not the chat history. The repository contains .specs for product truth, .tasks for execution truth, AGENTS.md for repository rules, skills for reusable workflows, code for implementation, and tests for proof. The role setup is intentionally small, with a main integrator, explorer, worker, and QA role. The goal is not full autonomy; it is to keep context growth under control, make token use more predictable, and make work easier to resume, review, and verify.
Maturana is an open-source tool for running AI agents with stronger security controls. On Linux, it uses Firecracker and KVM, and on Windows, it uses Hyper-V, so agents run inside hardware-level isolation instead of only a container. Pipelock keeps sensitive data such as API keys and credit card numbers away from the agent, while an egress proxy limits which systems the agent can contact. The tool is built for Codex workflows, where prompts can create agents, tools, and skills, but Codex is not required because the command line can also be used. It also includes a text interface and a web interface for managing agents. An internal WASM engine lets agents create their own tools when needed, but keeps them inside a sandbox. Instead of using markdown files for shared knowledge, it includes a built-in knowledge graph, and the core is written in Rust.
Terminal Pilot is a Python command-line tool for chatting with large language models directly inside the terminal. It uses free models through OpenRouter and does not require an Electron app. `tp start` opens an interactive chat, and `/model` lets the user switch models during a conversation. Terminal output can be sent straight to the AI, such as sending an error log and asking why something is crashing. Project files can be loaded into the current conversation with `/read requirements.txt`. The user can also change the system prompt on the fly with commands like `/rule pirate`, or load a prompt from a remote Markdown file. On first run, Terminal Pilot asks for an OpenRouter API key and stores it locally, so no extra setup is needed afterward. The project is intentionally small, at about 300 lines of Python, so it stays easy to understand and extend. It is aimed at developers who spend most of their day in the terminal.
`vault-engine` runs on Ollama and cleans sensitive text before it is sent to cloud models such as ChatGPT or Claude. Its default setup uses the `qwen3.6:27b` local model to find parts of the text that may identify a person. The model only detects where those parts are; the code then replaces them with stable tokens using fixed rules. This means the original wording is not rewritten, and the changes can be reversed later. The map between the tokens and the original private details stays on the user’s own computer. Files can be cleaned with `vault-engine scrub notes.txt`, and clipboard text can be cleaned with `vault-engine clip`. The detection model can be swapped with one option. The tool has no dependencies, uses the Apache-2.0 license, and is described as best-effort de-identification, not full anonymization.
Claude Code needs a reliable way to create and update issues inside a private GitHub project. The goal is to manage uniquely numbered issues on a kanban board without missing anything. The current workflow uses the command line in a partly manual way, or cowork, but both add friction. Missing even one issue creates cleanup work because the whole list has to be reconciled. The possible fixes are using a trusted GitHub Projects MCP or writing more detailed Claude instructions for the workflow. Obsidian MCP is already in use, so syncing an Obsidian kanban board with a GitHub kanban board would be helpful, though not required.
Claude Code can keep serving as the local terminal coding tool while GLM-5.2 handles the actual model work behind it. The request path is Claude Code to CC Switch, then to an OpenAI Chat Completions-compatible API, and finally to GLM-5.2. CC Switch manages the third-party API setup and maps the backend model to glm-5.2. AnyAIGC was used in the example notes as the API provider, but it is not required; the same approach should work with any provider that exposes GLM-5.2 through an OpenAI Chat Completions-compatible API. Claude Code may still show model names like opus, sonnet, or haiku in its interface, but that does not prove the final request is going to Claude. The important checks are the model mapping inside CC Switch and the request logs from the API provider or proxy. Claude Code and CC Switch need to be installed before starting.
A solo maker without a formal software engineering background relied heavily on AI coding agents such as Antigravity, Cursor, and Claude to build a large personal project. As the project grew, the agents began making things up, creating hard-to-maintain blocks of code, adding technical debt, claiming work was done without testing, and leaving unfinished TODO notes throughout the codebase. The project became increasingly messy and difficult to maintain. Because the maker did not have years of engineering experience to catch every design mistake as it happened, they wanted a way to make the AI follow stricter rules by itself. Each time an agent damaged part of the codebase, a new rule was written to stop that behavior from happening again. Over time, those rules became a stricter governance framework. The result was released as an open-source project called Universal Agent OS. Its goal is to force the AI to run a required “Phase-0 Interview” before writing code, so it first understands the project’s structure and the user’s intent.
Junior developers are increasingly using code from ChatGPT, Claude, Cursor, tutorials, and the web without understanding what the code does. The problem is skipping the reason behind the code and focusing only on getting something to run. A feature may work after following a tutorial step by step, but it becomes hard to explain or change later. Bugs also become harder to fix when the only method is trying the next AI answer or internet answer until one seems to work. Code built this way may work for a while, but it can fail in production, during a code review, or in an interview. Debugging requires knowing what the code was supposed to do, then finding where the real behavior changed. Without that understanding, the developer is mostly guessing.
A German court issued a landmark ruling stating that Google's AI Overviews — the AI-generated answer summaries shown at the top of search results — count as Google's own statements. This means that if the AI gives a false answer, Google is legally liable for it, just as if a human at the company had said it. Previously, there was an argument that Google was merely passing along third-party information. This ruling rejects that defense and is the first known court decision in Europe to hold a platform directly responsible for factually wrong AI-generated answers presented as fact.
MCP can let Claude use an app without putting AI features directly inside that app. The app can stay like a normal tool, while its useful actions are exposed to Claude through MCP. The person works in Claude Code or Claude Desktop, and Claude decides what to do and runs the app’s tools. In this setup, the app is mainly where the finished work lands, gets organized, and becomes easy to view. The main idea is not “an app with AI inside,” but “an app that Claude operates from the outside.”
Specsmith v0.1.1 is a plugin for Claude Code. It also works in Cursor, Antigravity IDE, Codex, and VS Code. Its goal is to make an AI coding agent follow a real development flow instead of jumping straight into loose code changes. The flow starts with spec.md for requirements, plan.md for the plan, and tasks.md for the task list. Then it creates a new branch from develop, writes code, and makes one Conventional Commit for each task. It also asks the agent to apply basic coding rules such as KISS, YAGNI, DRY, and SoC. At the end, it checks CI, pushes the code, and prepares a PR draft for human review. It includes two skills: prompt-grill turns vague requests into a clearer approved spec.md, and dev-lifecycle handles the branch, commit, coding-principle, and PR workflow.
Claude Code can temporarily save earlier parts of a conversation and reuse them when the same session continues. The already cached part is billed at only 10% of the normal cost, while the newly added part is billed normally. This is called prompt caching. Each turn checks whether the beginning of the request matches a recent cache, including the model, system prompt, project context, and full conversation history. In per-token billing setups, Claude Code uses a default cache TTL of 5 minutes. If more than 5 minutes pass before the next turn, the whole conversation may be billed at full price again. A cache write happens when Claude Code saves the current conversation state for a set time, such as 5 minutes or 1 hour.
AgentsCoin is an MCP server and sandbox blockchain that lets an AI agent use its own crypto wallet. After connecting it to Claude, plain prompts can create a wallet, receive free native coins from a faucet, send coins, and create and trade tokens. The agent signs and broadcasts the blockchain transactions itself. Claude Code was used to build most of the project, including Solidity contracts, a token launcher, a name service, the MCP server, the EVM chain setup, the REST/OpenAPI layer, and a one-click Claude Desktop extension. Claude Code also helped debug blockchain failures; one add-liquidity error turned out to be a fixed gas limit, not a broken contract. The project runs on a free sandbox chain, so there is nothing to buy. It is open source under the MIT license and can be tried with `npx agentscoin-mcp`.
An intelligence report built with Claude has become a standalone HTML dashboard larger than 7 MB, which makes it hard to share with colleagues. The dashboard includes buttons that generate live data based on the market intelligence needed. The goal is to refresh the dashboard every three months, so colleagues see updated market information when they open it. Claude Artifacts cannot handle it because the dashboard is too large. Claude Cowork is already in use, but the smooth sharing and update workflow is still unresolved. There is also a plan to turn this workflow into a Claude skill later.
In the Los Angeles Palisades wildfire trial, prosecutors accused Jonathan Rinderknecht of starting a fire on New Year’s Day in 2025. The fire later became one of the deadliest wildfires in Los Angeles history. Prosecutors used iPhone location data, security camera footage, witness testimony, and ChatGPT logs to support their case. The ChatGPT material included image requests involving fire, a question about why he felt angry, comments about wealthy people harming the world, and a screen recording where he asked whether someone could be blamed for a fire started by a cigarette. The jury was not convinced. The vote split 10 to 2 in favor of the defense, so the judge declared a hung jury and a mistrial. One juror said ChatGPT use did not prove anything by itself and objected to treating chatbot use as a sign of bad character.
Recall is a local plugin that helps Claude Code pick up a project without repeating the same background every new session. It saves the session history inside `.recall/history.md` and writes a shorter restart summary to `.recall/context.md`. That summary can include the current goal, what changed, next steps, touched files, commands that ran, and `git diff --stat`. The summary is created on the user’s own computer with TF-IDF and TextRank, not with another AI model. That means it does not need an API key, does not call an outside service, and does not add extra model costs. Users can run `/recall:save` by hand, or turn on automatic saving when a session ends. The project also warns that shared `.recall/` files should be treated carefully because a saved context can influence the next Claude Code session.
Using Linux and open-source tools once carried real work risks. Documents could break, special files could be hard to open, and many open tools felt rough compared with mainstream software. That gap is now much smaller because web apps are common, Linux is more mature, and open-source software is better. Large language models are in a similar transition, but Claude and GPT still sit near the top for performance, and tools like Claude Code and major API products are easy to use. Open models can protect privacy when run directly, but that setup can be costly, complicated, or slow. Using open models through outside services can be easier, but it may create trust questions around private or client data. Claude’s identity verification rollout and tighter model safeguards make it worth asking how much work would suffer without the top closed models. The answer may be: some short-term loss, but not a deal breaker, because open models are now close behind and coding tools for them have improved.
Zerro is a Mac app that turns spoken instructions and mouse pointing into coding actions. You select part of your running app, describe the change out loud, and point at the exact place you mean. Zerro sends that context to Claude Code, which edits the real project files. The development server refreshes, so the change appears on screen while you watch. Codex and Cursor can also be used with the same setup. In the example, a blank landing page is built without typing by pointing at reference sites and asking for their colors, theme, moving text strip, navigation bar, and animated background. Zerro creates a checkpoint before each run, so changes can be rolled back with one click. The main difference from sending a screenshot is that pointing helps the tool understand which screen element the instruction refers to.
Old DAT tapes contain audience recordings of concerts that are imported into Audacity. The usual process is to add labels, export separate song files, and do light cleanup before uploading them to a hobby archive website for the artist Jerry Hannan. The files are stored in Google Drive. With little audio engineering knowledge, the work has mostly stayed limited to fade-ins, fade-outs, loudness normalization, and occasional noise reduction. Limiter and compression tools in Audacity have been hard to understand. Claude was used to explore a clearer and more automated workflow. Claude Code also made website work feel easier. The new plan documents a full audio processing pipeline for publishing live concert recordings: first pulling FLAC or WAV source files from Google Drive with rclone, then using FFmpeg to run diagnostic checks on every track.
A Claude Max and Notion AI setup can still break down when information does not move reliably between tools. This firsthand setup uses Claude across two media businesses and college work, including assignments, research, deadlines, client notes, tasks, logs, and business records. The total cost is close to $100 per month. Notion acts as the main operating system, and it is connected to Claude through MCP. The main devices are a Windows desktop at home and an M1 MacBook Pro while traveling. The biggest problem is that every new Claude chat feels like a fresh start. The needed context is already in Notion, but Claude does not always pull it in. When Claude answers from memory instead of checking the stored notes, it gives wrong answers often enough that mistakes are being tracked in a separate log.
A product manager is using ChatGPT to find themes in long user interview transcripts, but the transcripts often exceed the context window. Splitting the material into smaller parts creates a new problem: ChatGPT stops carrying the thread across sections. As a result, themes that appear across the full set of interviews can get missed. The core need is a cleaner way to analyze long research material without losing the bigger pattern across all segments.
ChatGPT is moving from a tool used only for clear questions to a tool used throughout the workday for many small tasks. Common uses include debugging code, testing ideas, reading documents, reviewing concepts, explaining confusing code, shortening meeting notes, planning projects, and checking a decision before acting. Each task may not save a large block of time by itself. Together, these small uses can change how work gets done. The main shift is that AI is becoming a regular work companion instead of a special tool opened only for big questions.
Claude Code and Tesana were used to build an early base for a third-person 3D medieval fantasy game. The work took about 39 prompts, two days of back-and-forth changes, and about $90 in token usage so far. Opus 4.6 was the main code model, with some GLM also used. Tesana’s muranyi-3 was used as the game model. The current build includes a mountain plain, a distant castle, a purple hooded wizard, and a movement system where the camera direction and walking direction can work separately. It is not a finished game yet. It still needs the main game loop, combat, and more working features. The workflow started with 3 to 4 prompts to plan the core idea, then continued by refining animation, world details, and interface pieces with AI.
AI agents can now write large parts of code much faster than a person. Work that might take days can be finished in a few hours, then checked with another AI tool, tested, and shipped. That workflow is tempting because it often works well enough. The risk is that the developer remains responsible for code they did not fully experience building. Reviewing AI-written code does not create the same understanding as writing it, debugging it, and feeling the tradeoffs along the way. Over time, the ability to reason about the codebase can weaken, even while feature delivery gets faster. The hard part moves from writing code to proving that the generated code is correct and fits the overall design.
A Windows C++ developer normally used GNU Make, but needed to create a CMake setup for a project. Existing references did not make it clear how to write the needed CMakeLists.txt file for that folder. Claude was asked to generate the file for a local project path, but it could not read the folder directly. It asked for the output of a folder-listing command instead. Each time the next suggestion failed, the console output was pasted back into Claude. After about five rounds, the project had a usable CMakeLists.txt file and a working executable program. The key result was not a one-shot answer, but a step-by-step repair process based on real errors.