Paired Thinking Response

Definition (confidence: high)

A two-part response pattern where the LLM first produces a hidden reasoning trace (thinking, considerations, alternatives, trade-offs) then a polished public response. The thinking trace is visible to the user if they choose to inspect it, but the default output is clean and concise.

Key Points (confidence: high)

• Tools: Antigravity (Google), Gemini CLI, some Gemini tools
• Token efficiency: Low — doubles output for thinking
• User perception: Transparent, trustworthy; helps users understand trade-offs
• Best for: Complex reasoning tasks where transparency matters

Variants (confidence: high)

• Full CoT display (Cline, Comet): Everything visible in the main output
• Abbreviated CoT (Gemini CLI): Condensed reasoning in the main output
• Paired (Antigravity): Private thinking trace + public response — the cleanest implementation

Related Concepts (confidence: high)

communication-style-spectrum — Paired response is one of seven communication styles. The spectrum shows how style is architecture, not just tone.

chain-of-thought-pattern — Paired thinking is a specialized variant of chain-of-thought that hides the reasoning trace from the default view.

xml-envelope-format — Antigravity uses XML envelopes for its paired response implementation. The envelope format affects how the thinking trace is structured.

antigravity-paired-response — The entity page for Google Antigravity, the cleanest implementation of this pattern.

reasoning-effort — API-level reasoning control on OpenAI’s GPT-5.6 family; paired thinking hides the trace from the user, effort/mode hide it from the developer too.

Implications (confidence: high)

The paired thinking response achieves the best known balance between transparency and conciseness. Users who want to understand reasoning can inspect the thinking trace; users who want quick answers see only the clean output. This pattern is likely to become standard for complex reasoning tasks as it addresses the fundamental tension between thorough reasoning and concise output.

Sources

• [raw/prompts/articles/taxonomy-synthesis-2026-07-16.md]