What it does

Thinking Claude was a two-part hack: a lengthy system prompt that instructed Claude-3.5 Sonnet to narrate its reasoning process before answering, plus a Chrome extension that formatted that inner monologue into collapsible, copyable sections in the Claude.ai web UI. The author explicitly disclaimed any benchmark ambitions; the goal was simply to see how far “deep mindset” prompting could stretch daily-use responses.

The interesting bit

The project became a victim of its own success. Anthropic shipped native reasoning capabilities in newer Claude versions, and the maintainers promptly retired the repo with a refreshingly honest admission: the base model had absorbed the behavior they were jury-rigging. That kind of graceful obsolescence is rare in the prompt-engineering world, where abandoned repos usually just rot.

Key highlights

• Versioned instruction sets (v3.5 through v5.1-extensive) iterated on the thinking protocol over roughly three weeks
• Chrome extension v3.2.3 offered fold/unfold sections, one-click copying, and auto-formatting of new messages
• Firefox port was “in development”; legacy Chrome extension deprecated
• MIT licensed; 17k+ stars suggest strong curiosity about visible reasoning chains
• Manual installation required downloading and loading unpacked extension; Chrome Web Store version noted as potentially outdated

Caveats

• Retired: Repository explicitly abandoned as of the README’s CAUTION banner; no further development
• Firefox support never shipped: listed as in-development with no release evidence
• Setup friction: required pasting instruction files into Claude’s custom style UI, not a one-click install

Verdict

Worth studying as a case study in prompt engineering’s half-life—how a popular community workaround gets absorbed into base models. For current use, the native reasoning in modern Claude is strictly superior. Historians of LLM UI patterns and prompt-craft archivists only.