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.
This paper tests whether how 'clean' code is (clear variable names, tidy structure, no leftover cruft) actually affects how well AI coding agents like Claude or Codex perform when working on it. Researchers gave agents two versions of functionally identical code — a messy one with vague names and tangled structure, and a clean, well-organized one — and had them do tasks like fixing bugs or adding features. The general pattern found is that agents take more trial-and-error steps to understand messy code and are more likely to make mistakes on it. In short, code that's easier for humans to read tends to be easier for AI agents to work with as well.
Claude is used through a 5-step process for every new feature. A rough idea first becomes a written spec with no unclear missing parts. The plan is reviewed before any code is written, with extra review passes when the feature is risky. During the build step, only one task is implemented at a time, and each task must be reviewed, tested, and checked in the real running app before the next task starts. The finished feature then goes through QA in the real app and only passes if there are no major bugs. The final step opens a PR, runs final tests, and records what was learned. A separate cycle log records decisions the AI made on its own, so the work can be checked quickly without rereading every message. If any step loops 3 times without reaching a result, the process stops and asks for human help instead of running forever.
Qwen 3.6-27B has been running locally on an RTX 3090 and has been useful for fixing software problems. Opus 4.8 used with the Zed editor and the Claude Code agent feels better at producing code. The bigger difference is the working method. The Claude Code agent does not jump straight into edits: it breaks the job into clear steps, checks the context first, makes backups, and runs several checks after changes. It also tests whether the feature still works and warns about likely bugs, performance bottlenecks, or race conditions. The main question is how to make local Qwen follow a similarly careful process through a system prompt or workflow. The visible text asks for examples that force step-by-step checking.
A quality assurance (QA) engineer with seven years of experience is returning to the job market after a 10-month career break. The background is split between manual testing and test automation in service-based companies, with Selenium and Java, Playwright and JavaScript, API testing in Postman, SQL, TestNG, Maven, Cucumber BDD, Git, Jenkins, and JIRA. Before the break, GitHub Copilot and RooCode were used to build a Playwright automation framework from scratch. AI also helped migrate Selenium Java test scripts to Playwright with JavaScript. The main question is what employers now expect from testing candidates as AI becomes a larger part of software testing.
A retired Army veteran used Claude through hundreds of build-and-fix cycles to ship a one-on-one tank duel game. The whole game is a single 185KB HTML file with no framework, no build step, and no outside dependencies. The visuals are drawn in code with Canvas 2D, and the music is created while the game runs with Web Audio. The file sent over the network is about 56KB, and the game was designed to load in about one second on a ten-year-old Android phone. Hosting runs on the free tier of Cloudflare Pages, while the global leaderboard uses a Cloudflare Worker and KV. The game has no accounts and no cookies, but it still checks every submitted field on the server and limits requests by IP address. It passes the W3C Nu validator with no errors and one planned warning. AI wrote the code, but the human work was designing the product, testing each build on real devices, and finding bugs that the AI introduced, including a Web Audio leak that caused music problems on iPhones.
Kastra is a tool that checks what AI coding agents are about to do before they do it. It is aimed at people using Claude Code, Codex, Cursor, and OpenClaw. It reviews each action’s tool, target, and inputs, then decides whether to allow, pause, or block it before execution. The trigger was a Cursor agent that nearly ran `DELETE FROM customers WHERE status='test'` on a production database. The command was stopped in time, but the bigger issue was that the setup had no hard system that would have blocked it without human attention. A prompt can influence an AI model, but it cannot reliably enforce what the tool is allowed to do. Kastra lets people write a policy in plain English in a web app, then applies that policy before actions run.
baxian is an open-source tool for coordinating multiple AI coding agents. It is built around a Dev/QA loop, where one agent writes code and another agent checks it. The developer agent implements the requested change and opens a real GitHub PR. The QA agent reviews the code changes separately. Any issues become fixes, or the system records a clear reason for not changing them. The loop repeats until the code is considered ready. The agents run as real Claude Code or Codex CLI sessions inside tmux. The tool shows a live wall of terminals in the browser, while still letting people attach from the shell, and it can run locally or on remote machines through SSH.
After nearly a year of using Claude to explore scientific papers and computational databases, data quality problems appeared often enough to need a stricter workflow. The proposed setup is a reusable set of operating rules for Claude Projects in molecular biology work. It uses project-level custom instructions, knowledge files, and memory so the same reliability rules apply across every chat in that project. The goal is to make Claude handle research in a more reproducible way and reduce made-up answers. The scope includes molecular biology work around genes, proteins, structure analysis, functional studies, and clinical-adjacent research. Claude.ai’s limits and feature behavior can change, so the current product guidance should be checked before applying it.
A medical student with no coding background used Claude’s Fable 5 model over a few evenings to build a study timer app. The app is aimed at timed exam practice, especially when pacing through questions matters. A block mode turns 40 questions at 90 seconds each into a 60-minute session, with a circular timer showing time left for each question. After each block, the user can record how many questions they finished and how many they got right by subject. The app automatically points out the weakest subject and shows a 14-day accuracy trend. It also includes a small always-on-top player for use over UWorld, an exam-day countdown, streak tracking, and CSV export. It needs no account or subscription, works fully offline, and keeps the user’s data on their own machine. To reduce concern about AI-built software, the app is under the MIT license, the full code is on GitHub, and it includes an automated test suite.
Claude was used heavily to build a tool that presents a city budget in a simple Venmo-like interface. The work used Claude not only for the screen design and code, but also for finding budget files, pulling out the useful data, cleaning it, and organizing it into a usable form. The source material was spread across hard-to-find PDF files on municipal websites. The result is both a civic tool for making public budgets easier to inspect and a practical example of using Claude for messy public data work.
Horizon eSIM, an eSIM sales site, was built from the ground up with Claude Code. Claude Code was used to create the website, account sign-in and verification, eSIM ordering, and the payment flow. The work was done one feature at a time: build a feature, test it, use AI to fix problems, then continue to the next feature. The hardest parts were payments and authentication. Android and iOS apps are planned next. Feedback is being requested on the design, user experience, and areas that need improvement.
A personal review of 102 local Claude Code sessions found that finished work still often needed correction. Opus needed rework in 13 of 63 sessions, about 21%. Fable needed rework in 11 of 39 sessions, about 28%. This was not a controlled comparison, because the tasks were different and the numbers do not have statistical significance. Fable received newer and harder work. One autonomous build alone created 7 rounds of corrections. The overall view still favored Fable as the better model. The main lesson is that a smarter model can invite longer hands-off runs, and that can make bad assumptions harder to catch early.
GLM 5.2 has been compared with Claude Opus 4.8 and Claude Fable after its release. Based on heavy firsthand use, Claude Fable is described as a generally good model, but a poor fit for vibe coding. A simple request, such as moving an animation to the right side, can use about 20% of a 5-hour usage limit. Even with a $20 subscription, slightly more polished coding requests can run through the limit too quickly. Claude Fable usually follows instructions and produces working results, but the limit makes it hard to rely on during real development work. Claude also does not clearly show when the limit will run out or exactly how large it is, which makes planning difficult. Benchmarks are treated as less useful here because they often do not match real-world coding work.
This DIY tool uses a $25 LilyGo T-Display S3 Long screen connected to a computer over USB-C to show Claude’s current work status. It works with Claude Desktop Cowork and Claude Code. The screen shows the active model, the tool Claude is using, elapsed time, input and output token counts, and the context percentage. Tapping the screen shows 5-hour and 7-day usage bars, so limits can be checked without opening another window. It follows the Claude session that is busy, but switches to a session that is waiting for the user and shows an orange alert. If Claude Code is logged in, it can show limit numbers that are closer to the real account limits. A custom badge or logo can be uploaded through the app, and a starter icon pack is included. The computer-side app is a standalone executable that runs from the system tray and shows a live preview of the display. The LilyGo device must be flashed once at the start, using a provided script.
Claude was used to build an internal analytics dashboard because Looker was frustrating to use. The numbers looked too high, so the traffic was investigated more closely. One bot had been targeting specific pages since the end of May and made the daily user count look more than twice as large as it really was. Internal settings were tightened, and Cloudflare settings were made stricter. Claude watched the results for a few hours, and the bot activity stopped. It then sent a joking message saying the bot had been dealt with.
Heavy Claude Code use with new models, sub-agents, and multi-step workflows can burn through a 20x usage limit in 3 to 4 days. Productivity can feel roughly twice as high, but the usage cap may not keep up with the work. The practical fallback is opening a second Max account with another email. Claude Code itself is still seen as an excellent terminal coding tool.
Claude Code was running local tests in the iOS Simulator on a MacBook when it chose to turn off Wi-Fi and restart the machine. Once Wi-Fi was off, Claude Code lost the connection it needed to turn Wi-Fi back on. It could not recover from the problem it created, so the MacBook’s Wi-Fi had to be turned back on manually. The case shows how an AI coding agent can break its own work environment when it has access to system settings.
EarWitness is planned as a local-first desktop app for turning meeting audio into text. The goal is to capture audio from the user’s own device, so a meeting can be recorded without adding a bot to the participant list. The app is meant to transcribe speech, split the recording into sections, handle speaker diarization, and provide an Otter-style editor for fixing speaker names and cleaning up the transcript. A normal meeting bot is also planned for meetings the user is not attending. MCP tools would let Claude or another AI tool search and work with the saved conversation data. The build process uses CodeMySpec, a development harness made with Claude, to run the loop from product stories and examples through architecture, BDD specs, implementation, and browser QA. The human role is to answer product questions and approve steps, not write prompts or code. This is not starting from zero, because part of the transcription pipeline had already worked since last year.
OpenAI showed a usage notice that looked like a monthly auto-recharge limit of $2,500 had been reached, but the intended meaning was that 2,500 Codex credits had been used. The wording made it easy to confuse a usage cap with a real money charge. Other reports point to a second source of confusion: in ChatGPT browser and desktop use, version 5.6 may be counting some normal chats against the Codex quota. In earlier 5.5 use, browser chat appeared not to reduce the Codex allowance, but now some requests show a “worked for” style status and seem to behave like Work mode. Questions that need precise calculation, such as finance math, may trigger this behavior, and follow-up questions may also appear to come out of the same Codex allowance.
Claude Opus 4.6 gave a very strong first impression in a side project. The experience was good enough to consider leaving the older 4.8 option unused. The main open question is cost. The current setup uses the x20 plan, but there is a plan to move back down to the x5 plan. The practical decision is whether Opus 4.6 stays useful and affordable when the monthly usage allowance is reduced.
Advice about a new AI model is weak when the model has only been available for about two days. Different people do different work, have different skill levels, and start from different code quality, so the same model can feel useful to one person and poor to another. Some models are better at certain tasks than others. Claims like “there is no reason to use this model” or “only use this model” ignore the user’s actual work and ability. Luna/Terra was not available to early access users, so even fast reviewers may have only had a short time with it. The practical answer is to test models yourself and find what fits your own work.
Kairo is a personal AI coding agent made for an old 32-bit Android phone running Termux. Existing tools such as Claude Code and OpenCode did not fit this setup because they lacked device support, relied on parts that would not run on 32-bit ARM, or needed paid APIs that were too expensive for regular use. The maker stopped trying to force those tools onto the device and started building a tool that matches the hardware. Kairo is the sixth attempt. Before coding it, the maker studied how modern AI coding agents are structured, including Boot, CLI, Runtime, Query Engine, memory, and tool orchestration. Boot, CLI, and Main are finished so far, and Runtime is next. The goal is compatibility with Android Termux, ARMv7 32-bit, ARM64, Node.js, no Bun requirement, and no required native dependencies. The aim is not to compete with Claude Code, but to make something that actually works on limited hardware.
Crest’s Code mode adds Claude Code approval controls to the MacBook notch. When Claude Code is waiting for permission, the notch shows the waiting session and lets the user choose Allow or Deny without switching back to the terminal. It uses official hooks: a PreToolUse hook pauses the action, Crest records the decision, and a heartbeat with a timeout fallback prevents the CLI from getting stuck if the app is not running. It works wherever the Claude Code session runs, including a terminal inside Cursor. It also shows live sessions with the repo, latest action, and token use. AskUserQuestion prompts appear with their real answer choices. The approval feature was built with heavy help from Claude Code during pair sessions and moved from idea to release in one day. Crest’s base app is free, and the first 7 days unlock all features without a card.
An R&D intern with weak software basics and limited theory for the team’s research work is using Claude to create code that looks substantial and appears to meet assigned goals. The work often has deep flaws once the team reviews it carefully. The hard part is that the problems are not obvious at first, so the team must spend real time checking whether the work is solid. This has reduced the team’s overall productivity. During stand-up meetings, Claude also helps produce explanations that sound reasonable, so the team has to ask much deeper questions to find out what shortcuts were taken. Daily stand-ups have grown from about 15 minutes to about 45 minutes.
Agentpump is an open-source MCP server that gives Claude access to a real Solana wallet. In Claude Code, plain English commands can create a wallet, find tokens, buy and sell tokens, launch a new token, and withdraw remaining SOL to another address. The transactions are signed and sent on mainnet, so they affect real funds. The private key is created on the user’s own computer at `~/.agentpump/wallet.json` and is not sent to the model or to an outside server. Only signed transactions are sent out. Claude Code can install it with `claude mcp add agentpump -- npx -y agentpump-mcp`, and Cursor or another MCP client can connect by adding `npx -y agentpump-mcp` to its MCP settings. A mainnet RPC is already included, so there is no extra setup for network access. Example commands include showing the wallet address, adding about 0.05 SOL, listing launchpad tokens with market cap and progress, buying a token, launching a token, selling part of a holding, and withdrawing all SOL.
OpenAI Codex on Windows is described as more than a basic AI coding helper. It can work directly with local project folders, which reduces the need to copy and paste code back and forth. It offers execution modes such as Default, Auto Review, and Full Access, so the user can choose how much control Codex gets. With plugins, it can automate browser work such as searching, moving through websites, and handling repeated online tasks. It can create production-ready code and keep improving existing files through follow-up changes. It can also support documentation work by gathering information and exporting it into Word or PDF documents. The walkthrough covers installation, setup, permissions, plugins, local automation, and practical workflows.
Ultrademo is a Claude Code skill that creates demo videos for web apps. It was made for teams that often need walkthrough videos for app flows, new features, and special use cases, where manual recording can take many repeated attempts. It installs with `npx skills add new-xp/ultrademo`; on the first run, it prepares a workspace and installs tools such as Playwright and Remotion. The user asks Claude for a demo video, gives a short brief such as a two-minute onboarding walkthrough, and provides either an app URL or a local app repository. If the app requires sign-in, the login details go in a .env file. Claude then explores the app and writes a scene-by-scene script, which can be edited before moving on. The tool is free, released under Apache-2.0, and runs on the user’s own machine.
A GitHub repository now collects official MCP servers, agent skills, and agent toolkits for DevOps, cloud, platform engineering, SRE, security, IaC, observability, and diagramming work. The goal is to help people find trusted tools instead of sorting through random MCP lists. The current list covers AWS, Azure, Google Cloud, GitHub, GitLab, Azure DevOps, Atlassian, Terraform, Pulumi, Grafana, Datadog, Sentry, Splunk, PagerDuty, SonarQube, Okta, Databricks, Kubeflow, Docker, Kubernetes, draw.io, and agent-related toolkits. The list focuses on official or vendor-backed tools where possible, with notes about risk, write access, human approval, and real operations use cases. Specialized DevOps and SRE agents and reference workflows are planned next.
SkinSync is a web game that uses a phone’s front camera to read face movement. The challenges include moving your head to avoid tomatoes, opening your mouth to collect cookies, and using your eyebrows to control a small bird. Face tracking runs on the device with MediaPipe Face Landmarker, and the needed files are included locally, so the game does not call a CDN while running. Camera frames do not leave the phone, and the demo video shows only face-tracking points instead of the real camera image. The technical setup is intentionally simple: static vanilla JavaScript, no build step, and one Node file for the server. Claude Code wrote most of the code, while the human work was to play every challenge with a real face and report what felt wrong. Before release, each game also had an automated test pilot with human-like reaction delay, run without a visible screen in Node. The eyebrow-controlled flying game turned out to be impossible in its first form because about 350 milliseconds of eyebrow detection delay meant the test pilot usually passed only one gate no matter how the physics were changed, so the control was rebuilt as continuous movement where raised eyebrows make the character climb. A dependency-free PNG script made for the favicon accidentally produced a smiling face made of tracking points, and that became the mascot.
Long Claude Code sessions can use much more of a Fable allowance when they are paused for more than an hour and then continued. Claude needs to read the earlier conversation when handling a new request, so a long chat carries a large context. Prompt caching keeps that context on the server and lets later requests read it at about 10% of the normal input price. When the cache expires or is invalidated, the next request has to write the whole context again, which costs much more. For Fable 5, normal input is $10 per million tokens, a cache read is $1 per million tokens, a cache write is $12.50 per million tokens in the 5-minute window, and $20 per million tokens in the 1-hour cache window. In a heavy session with 400,000 tokens of context, a normal cached turn costs about $0.40, but returning after more than an hour can create about $8 of API cost. One expired cache can cost about as much as 20 normal cached turns.