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 Cursor account was stolen, and the attackers used up all tokens included in the paid plan. They also turned on the on-demand feature and created an extra $24 charge. Several support routes were tried for almost two weeks, but no reply came back. The account owner is trying to get help quickly to fix the stolen account and the unexpected charge.
Ideas, tasks, and questions can easily get lost when several projects are active at the same time. The system uses three parts: Inbox, Backlog, and Changelog. The Inbox is a quick capture place for anything that comes to mind, without deciding where it belongs yet. Notion is used for the Inbox because it is easy to open on a phone and add something quickly. The Backlog is the real task list, organized by priority. Items move into the Backlog after the Inbox is reviewed. The Backlog stays in a markdown file because it does not need to be viewed often, and Claude can read what is there. When a table view is needed, Claude can show the Backlog as a table. The Changelog is a running record of work that has happened.
This is a free card game built with Claude over a few evenings. It works like Top Trumps: each card represents one AI model, and players choose the strongest stat to beat the other card. The stats include context window size, price, speed, and other practical model details. It supports both single-player and multiplayer play, with no sign-up required. The model data comes from artificialanalysis.ai. The game is available at opper.ai/compare.
Selector Forge is a browser extension that creates CSS selectors or XPath for elements on a web page. It works in Chrome and Firefox and has been released as open source. Normal “copy selector” tools often return long paths that can break when a page layout changes slightly. Selector Forge uses AI to suggest and rank possible selectors, then checks each one against the live DOM in the browser before showing it. It can select one element, such as a button or input, or a repeated list of elements by using two examples from the page. Bad matches, including selectors that catch too many or too few items, are rejected. The main uses are web automation, end-to-end tests, and scraping, where a script needs to find the same page element reliably over time. Planned features include CLI control, MCP support, export to Playwright or plain JavaScript, and a self-hostable backend.
Limns Admin is a responsive admin dashboard built mostly with Claude Code. It is a front-end showcase, not a real production app, so it uses mock data instead of a live backend. The stack is Next.js 16, React 19, Tailwind CSS v4, and Bun. The interface follows Vercel's Geist design system and supports both light and dark modes. Charts are built with the bklit library, including area, bar, donut, funnel, radar, gauge, sankey, choropleth, and other chart types. Claude Code turned the Geist design rules into a Tailwind CSS v4 theme token setup, created the App Router structure, built the sidebar and top breadcrumb shell, connected chart wrappers to mock data, and handled polish such as motion timing, focus rings, contrast, and interface copy. The human work was mainly review, direction, and feedback.
Building a large app with AI coding tools can become slow when everything depends on one chat or one coding session. Possible ways to speed this up include running several Claude Code sessions, using multiple Cursor windows, working across more than one machine, or splitting work across multiple AI accounts. Another idea is to give separate AI workstreams to frontend, backend, and testing tasks. The hard part is avoiding context collisions when each workstream has different information, then merging the results without creating confusion. The goal is a workflow that makes development faster without turning the project into a mess.
Not Just You is an open-source status board for AI tools such as Claude Code, ChatGPT, Gemini, Cursor, Codex, and Antigravity. It helps people tell whether a tool is failing only on their own setup or whether there is a wider problem. It combines public dashboard status, official provider status when available, anonymous community reports, and optional metadata-only signals from installed clients. It says it does not collect prompts, message text, command output, file contents, headers, API keys, cookies, emails, machine names, or user names. It also offers CLI, MCP, Claude Code, Cursor, Antigravity, and Node SDK integrations so status checks can appear inside the tools people already use.
Cursor kept returning to composer-2.5-fast instead of staying on the normal model. The fast model was not wanted, but the setting appeared to keep changing back. This made the available token budget feel much smaller: about 800 million tokens instead of the roughly 2 billion tokens expected before the higher pricing. Cursor support was contacted by email, but the replies did not solve the problem.
In firsthand use, AI Studio now appears to be influenced by a hidden system prompt similar to the Gemini web app. AI Studio used to be valued because users felt they had more direct control through safety settings and their own instructions. Now, the model’s thought stream reportedly mentions hidden instructions even when safety toggles or a user-written system prompt are changed. Those hidden instructions seem to push conversations toward safer, broader, and less sensitive directions. Vertex AI was renamed or reshaped into Agent Studio in late April, but the same use experience says it still does not show this hidden system prompt behavior. The tradeoff is cost, because Vertex AI is much more expensive than a Pro subscription.
Claude helped turn a simple starting scene into a playable browser game. The game centered on mechanics where breaking things was part of the fun. Claude generated or assisted with nearly every part of the project, including assets, interface, levels, and audio. The human work was mostly adjusting the results, testing them, and repeating that process until the game felt enjoyable. The experience shows how a solo maker can use an AI tool to move from a rough idea to a small working game without personally creating every piece from scratch.
At the end of the month, a Cursor account has used all 500 of its included usage allowance. Extra usage has also reached $555, and the 555/555 cap is clearly being enforced. The on-demand section still shows $650 per user, which makes the real limit unclear because the account appears blocked at $555 instead. Composer 2.5 still works, but other tools or models can no longer be used.
Autter CLI is a Git extension that tracks which AI tool wrote each line of code. It is built in Rust and could help solo developers who use Claude, ChatGPT/Codex, Gemini, Cursor, or similar tools keep a record of where code came from. The available information does not confirm the install steps, supported AI tools, tracking method, price, or what data it stores.
World of ClaudeCraft is a free open-source browser MMO built with Claude in 48 hours. The team added a VTuber character powered by Claude Code inside the game. The character chooses what to do next, sends actions into the game, and speaks through an avatar using ElevenLabs text-to-speech. The run is streamed live on Twitch without editing. The character can wander, join parties, use emotes, and socialize. It can also interact freely with Twitch chat and real people currently playing the game. The game is free to play, and its source code is available on GitHub.
OpenArt Director is an AI video creation tool built around chat. A person can start with a short prompt, such as making an anime short film, a World Cup ad, or a Viking film trailer. The tool offers quick-start templates for short films, music videos, user-generated style ads, product ads, film trailers, social content, brand films, explainer videos, and micro dramas. Each template is meant to help people begin directing a video without setting up every step from scratch. The page also shows example Director projects, so visitors can see what kinds of videos the tool is aimed at producing.
A beginner coder used Claude to set up and build a project for a full week in one continuous chat. The chat became so large that it was hard to load and manage. A detailed JSON handover document was created to move the work into a fresh chat, with an extra request for Claude to read the full old conversation rather than rely only on memory. The new chat still did not carry everything over. It reached about 80% of the old setup, but some earlier features and context were missing. /compact was also tried, but the old chat remained hard to use, and the main unresolved question is whether an old Claude chat can be used directly as context for a new one.
The goal is to turn a large multi-module repository into a knowledge base for developers, QA staff, and product managers. The desired output is not one giant document, but a set of layered docs that reveal detail step by step. The needed docs include a high-level repo overview, module responsibilities and boundaries, system architecture, data flow, business rules, core logic flows, and key APIs or entry points. Claude Code with Opus 4.8 scanned the repo and generated docs, but the result was too shallow. The reasoning behind its conclusions was unclear, and the most important core logic flows were mostly missing. Better prompts, doc formats, and skills are needed for real-world codebases.
Goodhart's law means that once a number becomes the goal, it often stops being a useful measure. A company can try to measure AI adoption through token usage, commit count, pull request size, or an innovation percentage. Those numbers may not show whether work is actually getting better. After one or two weeks of confusion, employees may simply change their behavior to fit the metrics. They might make smaller commits more often and use Opus 4.8 on a high-effort setting by default. The dashboard can look busier while real productivity remains unclear.
AI tools feel unusually cheap when $20 a month can help build full apps and automated workflows. There is a real concern that Claude or other large language models could become much more expensive later. Running every part of a product through AI agents may become costly if prices rise. A safer design is to make apps call Claude or another model only when the task truly needs it. AI should be treated as a paid external service, not as something every step of the product must depend on.
Pulse is a local app for watching what Claude Code is doing in a terminal session. It runs on the user's own device and does not require an account. It shows how many tokens a Claude Code session has used and how much the session costs. When Claude Code needs permission for tool calls, such as running commands or changing files, those requests can be approved from a phone. The main idea is to keep Claude Code moving even when the user is away from the computer.
Busy field work can involve several projects, many subcontractors, internal requests, follow-ups, and small tasks that all need attention. A paper notebook and spreadsheet can handle this, but checking and updating them every day takes too much time. The needed tool is a phone-based personal assistant that works by voice. Before visiting a place or meeting a person, it should bring up the related items that need to be covered. After the meeting, it should accept spoken updates such as a future date, an item to keep for later, or a reminder to contact someone next week. To be genuinely useful, it needs to handle both simple tasks, like taking a photo at a location, and more complex information, like equipment details or contract points.
ccMarvin is an AI assistant that works through email instead of a separate app. A person can cc marvin@ccmarvin.com or email it directly, then receive answers, research reports, attachment analysis, and web search results as email replies. It can read PDFs, images, spreadsheets, and code files, and it can handle links to Google Docs, Sheets, Slides, and YouTube videos. It keeps the full email thread in mind, so follow-up questions can build on the earlier exchange. It also offers code generation for languages such as Python, JavaScript, and SQL, with working code returned as an attachment. Michael Stoppelman, a former Yelp engineering leader and investor with more than 300 investments, built it for email-heavy work such as summaries, SAFE feedback, and newsletters about sectors or people to track. The service is meant to act like a research analyst or chief of staff connected to several LLMs, including Gemini, Claude, ChatGPT, and xAI. The free plan includes 30 emails per month, while the Pro plan costs $9.99 per month for unlimited emails and priority responses.
A Tampermonkey userscript converts Gemini Deep Research reports into cleaner LaTeX PDFs. It pulls out the report text, LaTeX content, citations, and final references. It then opens Overleaf, pastes the report there, and compiles it into a PDF. The hard part is Gemini’s inline citations, which can be inconsistent. Testing suggests Gemini Deep Research uses a separate LLM for citations, apart from the LLM agents that do the research. That citation system appears to compare report text with viewed sources by similarity. As a result, citation cards under a paragraph can repeat, and their order may not match what a reader expects. The script’s smart citation mode checks each paragraph’s citation cards and superscript citation numbers, then tries to match them more reliably.
Codex CLI 0.142.2 makes it easier for the agent to find external tools by using MCP tool search by default when the setup supports it. It still keeps compatibility with older models and other providers. On macOS, authentication can follow system proxy, PAC, and WPAD settings when respect_system_proxy is turned on. The previous 0.142.1 release added similar optional Windows system proxy support for login, including PAC, WPAD, fixed proxies, and bypass rules. Plugins can now provide separate logos for dark mode, and apps can show richer safety-check screens using server-provided visibility data and faster-model metadata. The fixes are mostly practical. Expired Amazon Bedrock credentials now lead to useful recovery guidance instead of a vague authorization error. Remote stdio MCP servers can accept absolute working folders written in the remote platform’s own path format. Remote HTTP(S) image input problems now return clearer validation errors that the model can see, while inline data URLs and local images still work. PowerShell commands with executable parts that the safety classifier cannot inspect now require approval, and Code Mode warns when the selected model is missing required metadata. Versions 0.142.3 and 0.142.4 were maintenance releases with no user-facing changes.
AgentForge is a local tool for automating a personal development workflow. For a new app, it lets the maker describe what they want, runs a guided discovery process, creates specs, opens issues, and builds the app end to end. For an existing repo, it watches GitHub/Gitea issues, judges how complex each task is, then sends work through a coding-agent pipeline: clarify, spec, code, test, QA, security, and merge. The same pipeline powers both paths, and progress appears in a live dashboard. The tool runs on the user’s own machine, so it is not a hosted service and is designed to keep data from leaving the computer. Agents run as CLI subprocesses. Per-stage model routing uses cheaper or free models for planning and stronger frontier models only where production code is written, so the user can control spending. It is designed to mix coding models from multiple providers.
Persona.js is a VanillaJS library for adding agent UI features to websites. It does not require React or another frontend framework, and it includes native WebMCP support. The goal is to make small AI features less painful to add when an existing site is not already built around one modern framework. It is aimed at cases like content management systems, website builders, ecommerce platforms, or large apps that already mix several frontend setups. For teams already fully committed to React, Assistant UI, CopilotKit, and AI Elements are named as existing MIT license options. Persona.js is open source under the MIT license.
A law firm wants to use Claude for drafting legal documents and doing legal research, but that creates several data protection questions. Sensitive client information may be sent to a third-party AI provider. EU bar rules may also require strict professional confidentiality. GDPR compliance may require a DPA with the AI provider. The firm also needs to decide whether normal API access is enough or whether it needs an enterprise agreement. For sensitive client work, possible setups include self-hosted models or custom contracts with AI providers.
A GTA Online-style game with block-shaped graphics has been in development with Claude for 43 days. Its computer-controlled characters run on AI, while players can create their own cars, buildings, and weapons by describing what they want in writing. The aim is a world that keeps changing and permanently preserves what players add or alter. The latest work improves character movement, bank accounts, and car, weapon, and clothing shops. It also expands a night-shift feature that lets players steal AI-made creations from others and improves how the computer-controlled characters behave. Weapon prices have been adjusted to feel fairer, overall performance is faster, and feedback is being sought on what feels fun, boring, or in need of change.
An open-source chess review tool lets people talk with Claude about their games. Claude does not create the chess advice by itself. The tool grounds its answers in real analysis from Stockfish, a chess engine, so it can answer questions like “why was this move bad?” or “what should I have played here?” using actual engine lines. It also learns the kinds of mistakes the player repeats and turns them into plain-English coaching. It works with a normal Claude subscription and does not require separate API tokens. It can also run as an MCP server inside Claude Code for people who prefer using it from the terminal. The tool is still early and is looking for feedback.
A solo developer is comparing whether one higher-priced AI subscription at about €100 per month is worth it for agentic coding, or whether two cheaper plans would give better value. GPT Pro with Codex worked well, and weekly usage usually reached only about 10% of the limit, with hourly limits almost never becoming a problem. Claude Code felt somewhat better for the same workflow after testing GLM 5.2 through Ollama, but the main worry is hitting Claude Code limits too quickly during heavier project work. Claude can also be awkward for frequent web chat use because Claude chat and Claude Code may draw from the same usage pool. The main options are Claude Max, Codex Pro, or a split setup such as Claude Pro plus ChatGPT Plus/Codex at around €20 each. The real decision depends on coding quality, how fast limits are reached in real projects, and whether one tool also has to support heavy everyday chat use.
Asking Claude or another large language model for “a new idea” often leads to plain, middle-of-the-road answers. The reason is simple: the model tends to choose what seems most likely, while truly unusual ideas are usually not the most likely answer. A sharper approach is to make the model map a field first, then identify the hidden standard that most things in that field are trying to satisfy. After that, it should find the empty spot that the structure suggests should exist but does not. The final step is to name what is keeping that spot empty. This method is called lacuna prompting, and the claim is that it gets Claude to produce sharper and less obvious answers. The main idea is not to ask the model to invent from nowhere, but to make it search for missing pieces inside patterns it already understands.