# Thinking-Claude: a retired prompt protocol and its Chrome extension > Thinking-Claude is a set of Claude instruction files plus a Chrome extension that folds the model's thinking into collapsible sections. The repository is retired, so the useful question is what it still does and what replaced it. **richards199999/Thinking-Claude** — Let your Claude able to think - Repository: https://github.com/richards199999/Thinking-Claude - Stars: 17,026 · Forks: 1,963 - Language: TypeScript - License: MIT - Published: 2026-09-21 · Updated: 2026-09-21 · Language: en - Canonical page: https://hysenlabs.com/projects/richards199999-thinking-claude ## What Thinking-Claude solved before Claude had native thinking The project targets a narrow problem: Claude 3.5 Sonnet in the Claude Web App answered quickly, and the answer arrived without visible deliberation. Thinking-Claude is two things bundled under one name. The first is a Thinking Protocol, a long instruction text stored in model_instructions/ and pasted into a Claude custom style. The second is a browser extension that reformats whatever reasoning text Claude produces into collapsible blocks. The README is explicit about scope. It says the project "is not aimed for benchmarks or huge leaps in math or something", because those outcomes are "pre-determined by the base model". The stated goal is exploration of Claude's "deep mindset" and the claim that the inner monolog is "very very fun and interesting" in daily tasks. That framing matters when you judge it. This was never sold as a capability upgrade. The audience was individual Claude.ai users, on either the Free or Pro tier, working in the browser rather than the API. Nothing in the repository layout suggests a server component, a proxy or a key store. The extension reads the page; the protocol changes what the model is asked to do. ## How the instruction files and the extension work together The two halves are independent, and that is the most important architectural fact. The extension does not inject the protocol. It only renders output. You can paste an instruction set and get thinking text with no extension installed; you can install the extension and see nothing extra if the model was not asked to think in that format. The protocol side is versioned prose. The repository tree lists model_instructions/v5.1-extensive-20241201.md, v5.1-20241125.md, v5-lite-20241124.md, v4-20241118.md, v4-lite-20241118.md and v3.5-20241113.md, with a changelog.md alongside. Each file is a self-contained instruction block, so switching versions means replacing pasted text, not running a migration. The extension side lives under extensions/, split into chrome/, a deprecated chrome_v0/ legacy build, and a firefox/ directory that the README describes as in development. The Chrome extension is the shipped artifact. Per the README it "automatically organizes Claude's thoughts into neat, collapsible sections", folds and unfolds parts of the reasoning, copies any part with one click, and reacts to new messages as they arrive. The mechanism is DOM-level: the extension watches the conversation and restructures rendered text. That is why it is tied to claude.ai's markup and why a Firefox build would be separate work rather than a repackage. ## Installing the Thinking-Claude Chrome extension and a first run The README recommends manual installation over the store listing, which it marks as possibly outdated. Download the v3.2.3 zip from the Releases page, unzip it, then load the unpacked dist folder. The store route exists at the Chrome Web Store under the slug thinking-claude, but the README's own caution is the reason to prefer the zip. ```bash # after unzipping the release archive # open chrome://extensions/ in Chrome # enable "Developer mode" (top right) # click "Load unpacked" and select the unzipped folder: dist ``` What you should see is the extension listed as loaded from your local dist directory rather than from the store. If the folder you select is the archive root instead of dist, Chrome will not find a manifest and the load will fail silently in the list. The protocol is pasted, not installed. The README walks through Claude.ai: open the style selector at the bottom of the input box, choose Create & Edit Styles, then Create Custom Style, then Describe style manually, then Start from scratch, then Use custom instructions (advanced), and paste the contents of the model_instructions file you want. ```bash # source file to copy from, for example: # model_instructions/v5.1-extensive-20241201.md ``` After that, start a chat. The README says the extension then makes the thinking process readable on its own. Expect the reasoning to appear in folded sections above the answer, with copy controls on each part. ## The retirement notice is the real limitation The README opens with a caution block stating that "This repository has been retired" and gives the reason: Claude is already a reasoning model with strong built-in thinking capabilities, making the project no longer necessary. That is an unusually direct self-assessment, and it should govern your decision. The practical consequences are concrete. The newest release listed is chrome-extension-v3.2.3 from 2025-01-09, and the last push to the repository was on 2026-04-07, so the extension is not tracking changes to the Claude.ai interface. A DOM-rewriting extension depends entirely on markup it does not control. When claude.ai changes its message rendering, the folding logic is the first thing to break, and there is no maintainer commitment to fix it. The protocol side has a different failure mode. Pasting a long instruction set consumes context and shapes every reply in that style, including short factual questions where a structured thinking routine adds latency and noise. The README's own framing, that the inner monolog is fun to read, hints at the cost: you are trading tokens and response time for visibility. If you want the model's reasoning to be reliable rather than entertaining, a hand-written prompt is a weak instrument compared with the model's own reasoning controls. It is also the wrong tool outside the browser. The project is built around the Claude Web App and a Chrome extension. There is no documented API integration here, and nothing in the repository layout suggests one. ## What to use instead, and where the difference actually lies The alternative the README itself points at is Claude's built-in thinking. That is not a competing repository; it is the model behaviour that made this one redundant. The difference in approach is fundamental. Thinking-Claude asks the model, in prose, to follow a thinking routine and then relies on a browser extension to make the resulting text legible. Native reasoning is produced by the model as part of generation, so there is no instruction block to paste, no version file to swap, and no DOM rewriting that can fall out of sync with a page redesign. If you specifically want the folding and copy behaviour, the honest comparison is against doing nothing. Claude.ai already renders its reasoning in a collapsible region, so the extension's main value was in an era when the raw thinking text arrived as an undifferentiated block. On a current interface that value is largely duplicated. For API work, the relevant comparison is with the reasoning controls exposed by the API rather than with any community prompt. The README does not document an API path for this project, so there is no migration story to describe here beyond dropping the prompt files. ## Maintenance cost, licence and what a fork inherits The repository is MIT licensed, which permits use and modification. The README states this plainly and adds "feel free to use and modify as needed". For a retired project that is the useful part: the instruction files and the extension source remain available to read, adapt and reuse. Upgrade cost is where the picture turns. There is no dependency graph to update in the usual sense. The root package.json is private, declares no runtime dependencies, and its only script is "prepare": "husky" with husky ^9.1.7 as the sole devDependency. So the root package manages Git hooks, not the extension build. The real upgrade burden sits in extensions/, and the versioned instruction files mean any change to the protocol is a manual re-paste for every user who adopted it. A fork inherits two liabilities. First, the extension must be kept aligned with claude.ai's DOM, and the README offers no compatibility contract. Second, the Firefox directory is described as in development, so a cross-browser fork is unfinished work, not a configuration flag. Anyone reviving this should treat the extension as the maintenance surface and the prompt files as frozen content. ## Conclusion Thinking-Claude is worth reading if you want to see how far a hand-written thinking protocol could push Claude 3.5 Sonnet, or if you need the extension's folding and copy behaviour on an older setup. It is the wrong choice for anyone starting today: the README states the repository is retired because Claude now reasons on its own, and the last push was on 2026-04-07 with the newest release from 2025-01-09. Before adopting anything, open model_instructions/ and confirm which version file you would paste, then check whether the Chrome Web Store listing still matches the v3.2.3 zip, because the README marks the store route as possibly outdated. ## FAQ ### What does thinking do on Claude? In this project, thinking is the visible reasoning text the model produces before its answer, which the Thinking Protocol asks Claude to write out. The README states the repository has been retired because Claude now has strong built-in thinking capabilities. ### How to show thinking in Claude? Paste one of the instruction files from model_instructions/ into a Claude custom style, then install the Thinking-Claude Chrome extension, which the README says organizes Claude's thoughts into collapsible sections automatically. ### How to tell if Claude is thinking? With this project, the thinking appears as folded sections above the response, and the extension lets you unfold any part of it. The README describes the extension as working automatically with new messages. ### Can you talk to Claude while it is thinking? The README does not address this. It only describes the extension reacting to new messages and folding the thinking that Claude produces before answering. ### How to use extended thinking Claude? The README does not document an extended thinking feature. Thinking-Claude's own mechanism is a pasted instruction protocol plus a Chrome extension that renders the resulting reasoning text. ### What is Thinking-Claude? It is a project with two parts: a Thinking Protocol made of instruction files under model_instructions/, and a Chrome extension that organizes Claude's thinking into collapsible sections. The README states the repository has been retired. ## Sources - [Issues](https://github.com/richards199999/Thinking-Claude/issues) - [License: MIT](https://github.com/richards199999/Thinking-Claude/blob/main/LICENSE) - [README](https://github.com/richards199999/Thinking-Claude/blob/main/README.md) - [Releases](https://github.com/richards199999/Thinking-Claude/releases) - [richards199999/Thinking-Claude on GitHub](https://github.com/richards199999/Thinking-Claude) --- Hysen Labs editorial analysis, written from the project's own repository and release notes. Cite the canonical page: https://hysenlabs.com/projects/richards199999-thinking-claude