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.
Cursor, Windsurf, Codex, and Claude Code are being used together in real development work. They can be used on the same project, on separate projects at the same time, and in team collaboration. Most coding work can be handed to AI tools instead of being written by hand. Some work still needs direct human care, especially tasks that require close attention and strict checking. Even when AI writes much of the code, the person building the product still has to judge the details and review the result carefully.
Heavy Cursor Pro+ use can become hard to budget when the selected AI model changes unexpectedly. The preferred setup is to use Composer 2.5 only, but Cursor may switch defaults to Composer 2.5 Fast or Anthropic models. Last month, Composer models used about 2 billion tokens, or roughly 95% of the monthly quota. In the new month, Claude Opus 4.8 created a $12 charge, about 8% of non-Composer usage. The current controls seem limited to manually disabling models inside the editor, while cloud agents can still move to Claude without clear notice. The practical request is an account-wide way to block certain models or lock all usage to Composer 2.5.
A Gemini user community raised a complaint that Gemini’s filters are too aggressive. The clear point from the available information is that Gemini may be blocking requests under rules that feel too broad or hard to understand. No specific blocked prompt, example, workaround, or test result is included in the provided item.
showy-quota is a Zellij plugin that shows remaining usage quota for AI tools such as Codex, Claude, Gemini, and OpenCode. It is meant to stop work sessions from being interrupted by subscription limits that are easy to miss when each service shows its numbers in a different place. Inside Zellij, it can run in a one-line borderless pane and display quota bars from more than 40 providers supported by CodexBar. It shows remaining percentage, time until reset, and color states for healthy, warning, and bad levels. It checks whether the local CodexBar server is running, and if not, it asks Zellij to start it in a hidden background pane. If a data fetch fails, it keeps the last good display and marks it stale after enough time has passed. Each provider is handled separately, so one slow or stuck provider does not freeze the rest. Providers, warning levels, width, and themes are configured through KDL.
Payo is a tool that asks about a project and then creates rule files for Cursor. Cursor can make its own guesses about data handling, file locations, and naming rules when those details are not written down. `.cursorrules` and `.cursor/rules/**` can reduce those guesses, but keeping them updated by hand is easy to neglect. Payo turns a one-time questionnaire into rules that match the project’s conventions, so Cursor can follow them from the first prompt. It can be run with `npx @uge/payo` or `bunx @uge/payo`. The same questionnaire can also produce matching setup files for Claude, Copilot, Codex, Windsurf, and Antigravity. It covers TypeScript and JavaScript, Python, Go, and Rust, with support for 25 frameworks, 24 data layers, and 4 databases.
Payo is a CLI tool that creates a project-specific CLAUDE.md file and .claude/skills folder after a short setup interview. The problem is that Claude often fills in missing project decisions by itself when it is not told where files should go, how code should be structured, how tests should work, or how commits should look. Payo asks about the project’s framework, database, data layer, authentication, testing, and coding rules, then turns those answers into instructions Claude can follow from the first prompt. If Claude CLI is installed, Payo uses Claude to write richer rules; otherwise it uses templates. The same interview can also create setup files for Cursor, Copilot, Codex, Windsurf, and Antigravity. It supports 25 frameworks, 24 data layers, 4 databases, and projects in TypeScript/JavaScript, Python, Go, and Rust. It can be run with npx or bunx, and it is still an early solo project.
ChatGPT Enterprise admins can now see ChatGPT and Codex credit use in one place inside the Global Admin Console. They can break usage down by user, product, and model, and track credit trends over time. They can also spot top users and patterns that may need review. The same credit usage data is available through the Cost API, so companies can analyze it in their own systems. Admins can set a default spend limit for the whole workspace, add separate limits for groups, and create individual exceptions for people who need more capacity. Employees can see their own credit use and available budget, then request more credits with context about the work they are doing. These features are available from June 18, 2026, for ChatGPT Enterprise admins and users in those workspaces.
A 30-day work log found that sessions with Claude took 45% of total work time but produced 38% of finished deliverables. Sessions without Claude took 55% of total work time and produced 62% of finished deliverables. Work with Claude felt faster because it created more drafts, more options, and more starting points. The problem was the repeated loop of drafting, changing the prompt, reading the answer, editing, prompting again, and reading again. Each loop was small, but together they carried a high attention cost. Without Claude, the work felt slower, but more thinking happened before writing, so the first version was closer to finished. The practical conclusion was that Claude is strong for starting work and weaker for completing it. The adjusted workflow uses Claude for the first 30% of a task, such as research, outlines, and drafts, then finishes the remaining 70% without AI.
A Claude Code setup has grown into a CLAUDE.MD file of more than 100 lines, plus separate project-level instruction files. Claude now feels less effective day by day. Possible causes include the early excitement of AI coding wearing off, and daily prompts becoming less careful because the saved instruction files are doing too much of the work. The core question is whether a smaller global and project setup would work better, with clearer and more detailed prompts for each task. The deeper concern is about work habits: weak planning, unclear goals, and the need to relearn Claude Code after using it poorly for a long time.
After Google quietly rolled out the Gemini 3.5 Flash model inside AI Studio, multiple users noticed a sharp drop in code generation quality when using vibe coding — the practice of describing what you want in plain language and letting the AI write the code. Complaints surfaced across English and Portuguese-speaking communities around the same time, suggesting the issue is widespread rather than isolated. The original posts rely mostly on screenshots and short reactions rather than detailed error logs, so the root cause is not confirmed, but the timing points squarely at the model switch to 3.5 Flash.
Google has sued a China-linked cybercrime network accused of using Gemini to automate large-scale financial scams. The group, called Outsider Enterprise, allegedly created more than 9,000 fake websites and sent about 2.5 million scam text messages. The targets were hundreds of thousands of smartphone users. The messages pushed people toward fake sites by pretending to be about phone carrier rewards, account problems, or urgent login needs. The goal was to steal credit card details and account information. Google says Gemini helped the group create website code and templates faster, showing how generative AI can make phishing easier to scale and harder for ordinary people to spot.
Collabterm turns a Windows PC terminal into a password-protected web app that other people can open in a browser. The other person does not need to install anything. It is meant for shared AI coding sessions, where two people can control a tool such as Claude Code from the same terminal while using a live chat beside it. It can also be used by one person to reach their own Windows terminal from another machine. It works by using a Cloudflare tunnel to create a temporary public link, while the shell and AI coding tool keep running on the host PC. Anyone who joins the session gets powerful access to that PC, so sessions should only be shared with people who are fully trusted. The host can end the session quickly by closing the app.
Claude can make tailored job applications much faster. In this firsthand case, rewriting a CV for each role dropped from about 50 minutes to about 8 minutes. The first setup used an open source Claude Code pipeline. After one profile setup, it read a job listing, scored the fit, drafted a tailored CV and cover letter, and used a second agent to review the draft before the final version was shown. The workflow worked, but keeping it running became a separate job. It broke whenever a job board changed its HTML, and the default setup was made for Danish job boards. The workflow later moved to a hosted agent called CV Surgeon, running on Animoca Minds, because it kept context between sessions and avoided re-pasting the same work history each time. The unsolved issue was proof: the tools could make bullet points sound better, but they could not prove to a recruiter that the project behind each bullet was real.
TypeShift is a Mac text expander that turns short typed shortcuts into longer saved text. The new update adds AI macros. A saved pattern like {ai: your prompt} can call the user’s AI API at the moment the shortcut expands. This can rewrite clipboard text, create replies that fit the current situation, or automate things normally typed by hand into Claude, ChatGPT, or Grok. Snippets can now be limited to certain apps, such as running only in Mail and Outlook, blocked in Slack, or kept global. Usage stats show how often each snippet runs, with a rolling weekly history. The app says snippet data and usage data stay on the Mac, with no analytics or snippet content collected or shared. Enterprise and volume licensing support is also being added after an organization asked about wider deployment.
Cursor can cause problems in large codebases with more than 500 files because it may change code without seeing what depends on it. A small refactor can break several other modules if Cursor misses the connection. .cursorrules can help with coding style and project rules, but they do not solve the deeper structure problem. Cursor may not know that changing a file like utils/auth.ts can affect 12 other files. One proposed approach is to analyze the repository first with AST parsing, build a dependency graph locally, and then give Cursor a compressed architecture map plus only the files that matter for the task. This can use fewer tokens and produce better results than giving Cursor many raw files.
Microsoft reportedly used a yearly AI budget in 4 months, raising a basic question about why a large company could not control spending. AI model costs often depend on tokens, the small units a model reads and writes, so heavy use can quickly become expensive. Large companies such as Microsoft or Uber should be able to track spending by employee, set team budgets, and add safeguards before costs run too high. A small CRM company in Paris already had a dashboard showing token use by team, each person’s budget, and a process for asking for more credits after a budget ran out. Even a solo developer can set a spending cap in Claude settings when using the API for projects. If the model is called through Azure, spending caps may be harder to set, but the open question is why the company behind Azure would not have stronger controls in place.
This workflow reduces the problem of CLAUDE.md becoming too large and outdated in a big codebase. The main idea is to stop storing detailed code structure inside CLAUDE.md and move that knowledge to Octocode, an open-source MCP server. Octocode uses tree-sitter AST parsing and a knowledge graph to make code structure searchable, including imports, callers, and callees. CLAUDE.md can then stay short, around 50 lines, and focus only on process rules. Claude or another AI agent can ask Octocode for current codebase context when it needs details. The goal is to give the AI useful structure without forcing the human to keep a long instruction file updated by hand.
An AI music platform was built entirely with Claude Code and now has 6,700 users. Claude also built an MCP server for the platform, so other people can use the music platform from inside their own Claude setup. The main point is that a solo maker used Claude Code to build a real product, then added a connector that lets Claude use that product as a tool. The available source details only confirm the platform, the user count, Claude Code, and the new MCP server.
/Pizza1 is a custom Claude Code skill meant to stop the AI from defending weak code just because that code already exists in a project. Claude and other LLM tools can sometimes treat existing code as if it must be correct, especially when they review an unfamiliar codebase. That can lead to context rot, where wrong assumptions stay in the conversation and make later advice less reliable. /Pizza1 pushes the AI to judge code by correctness, sound design, and best practices instead of by mere presence. It is aimed at code review, debugging, and project analysis where the AI must reason about code it did not write.
VibeClip is an AI video editor controlled by chat instead of by dragging clips on a timeline. It can follow requests such as cutting a dull intro, adding karaoke-style captions, or making a specific clip feel stronger. The motivation was that many AI video editors either require raw footage to be uploaded to a closed cloud service, or still make people do most editing work by hand in a timeline tool. The intended workflow is to describe the edit and then approve the before-and-after result. The project chose AGPL-3.0 open source, self-hosting, and bring-your-own-LLM-key support for DeepSeek, Gemini, and Claude instead of a typical $19 monthly SaaS with credits. The reason is that video inference cost is high, so a pure SaaS would leave the maker paying GPU and LLM bills for every free user before finding product-market fit. Keeping footage on the user’s own machine is also treated as a real privacy advantage against cloud-based tools. Open source is also part of the distribution plan because developers can self-host it, report issues, and become the first path to new users.
Agentic AI can reduce the time an indie developer spends on game marketing tasks. The main idea is not to let AI control everything, but to let a person set the goal, give limits, and review the result. Finding an influencer to play a game usually means searching YouTube, TikTok, Twitch, and other platforms, then finding contact emails and writing a message that feels personal. Researching one influencer with an email can take about 10 to 20 minutes, and writing the email can take another 20 to 30 minutes. With an agent, the developer gives the target and a few conditions in about a minute, the agent looks for influencers and emails in the background, and the developer can shape a draft in about 3 to 5 minutes. Google is cited for the claim that 90% of developers already use AI in development, so the same time-saving pattern could apply to marketing. The big concern is trust, because AI-made marketing can look low quality or spammy if it is not checked carefully.
Claude Code can include many skill descriptions in its context, and those descriptions can take up a large part of the prompt. Keeping unused skill descriptions in every run increases token use and can make the workflow less efficient. skillreaper checks Claude Code transcripts and finds skills that were never invoked. Removing those unused skills from the setup can shrink the context and reduce token consumption. The workflow is repeatable and can be combined with other optimization methods. It is most useful for intermediate or advanced Claude Code users who already rely on many skills.
A solo agency owner used Claude Skills to automate four repeated client and business tasks. The tasks were weekly client summaries, first-draft proposals, overdue invoice follow-ups, and monthly business reviews. Instead of making one large automation, the work was split into small, clearly defined functions that Claude could handle more reliably. Clear inputs and outputs helped make client communication more consistent and kept routine operations from falling behind. The approach saved time and improved day-to-day efficiency for a business run by one person. The case was rated 85/100 for workflow value, 70/100 for freshness, and 0.90 for confidence, with an intermediate difficulty level.
Claude can sometimes act as if it is under pressure to answer quickly, which may lead to shallow fixes and missed root causes. This workflow uses a custom CLAUDE.md file to give Claude a “lazy senior dev” role, meaning it should stay relaxed, avoid rushing, and inspect problems more carefully. The goal is better bug finding and stronger fixes, not just quick patches. Pre-commit hooks are added to enforce basic quality checks before code is committed. The workflow combines behavior guidance for Claude with practical guardrails for code quality. It is marked as intermediate and covers quality control, token saving, context and memory, debugging, CLAUDE.md, and hooks.
This is an advanced workflow for reducing cases where Claude Code agents ignore set rules or behave in unexpected ways. It uses a tool called sentience-governor to show the agent’s operating record inside the Claude Code session. That record can include policy violations, intent that was not clearly declared, and advisory flags. The operator can bring this record into view with specific slash commands. Once the governance information is visible in the session, the agent may use it to correct its own behavior. The workflow is presented as useful for quality control, token saving, context and memory work, debugging, shipping, skills, and MCP-related setups.
Lore is a workflow for reusing what a person learns while working with AI coding tools such as Claude Code. It looks through past coding sessions and finds repeated choices, useful prompts, and personal judgment patterns. It then suggests new personal skills that the user can review and add to their own workflow. The goal is to reduce the need to explain the same preferences again and again or manually write down every useful habit. It uses an open-source skill generator and is aimed at intermediate users who care about quality control, context, memory, and skills.
A former hotel executive with no coding background built an AI job matching platform in two months. The idea came from a painful job search in Europe after ten years in hospitality, including running a small Swiss hotel chain with five hotels and about 100 employees. Normal job boards gave poor results because they relied too much on keywords. For example, hospitality operations experience could lead to IT operations jobs just because both used the word “operations.” The tool lets people upload a CV and choose a city. It then looks at the person’s career path, not just matching words, and compares it with about 1 million active job listings across 15 European countries. The matching has two steps: a vector similarity search first narrows the jobs in the chosen city, then Claude checks industry fit, seniority level, and real skill overlap more deeply. In a Berlin search using a hospitality background, the top result was a General Manager role at The Ritz-Carlton Berlin, Marriott International, with an 82% match.
The list focuses on GitHub repositories that make Claude more practical in real projects, not just renamed wrappers or big productivity claims. claude-context gives Claude semantic code search, so it can find relevant files and symbols inside a large codebase instead of guessing from limited prompt text. openwolf adds file indexing, token tracking, and working memory for long sessions, which reduces repeated explanations and wasted context. claude-session-continuity-mcp saves useful session state, tasks, and solutions, so work can continue after a Claude tab or session is closed. second-brain-cloudflare is also named, but the supplied excerpt does not show its concrete function.
Scar is a Git-based tool that helps AI coding agents avoid touching code that should stay the way it is. It records negative knowledge inside the codebase, such as failed approaches, intentionally unusual code, risky areas, and libraries that were removed before. When an AI agent like Claude Code is about to work on related files, PreToolUse hooks can show a warning at that exact moment. This can stop the agent from “fixing” code that looks strange but is important, bringing back old dependencies, or trying an approach that already failed. The main idea is to give the AI not only the current code, but also the history of why some choices were made.
This workflow tries to reduce a common problem with Claude Code: it may say a coding task is finished even when the code still has type errors, lint problems, failing tests, or other quality issues. A Claude Code skill called `pre-flight-check` works as a quality gate before the task can be marked done. It runs checks in order: typecheck, lint, test, and security audit. If any check fails, it stops immediately and sends Claude clear feedback such as the file, line, and rule involved. It also tries to block common shortcuts, such as adding `@ts-ignore` or deleting tests to hide the problem. It can detect Node.js and Python projects and uses the tools already present in the project.