What it does This repo is a usage guide, not code. It documents two special control instructions that append to the first user message in DeepSeek-V4, switching how the model’s visible reasoning (the tischer / thinking block) behaves during roleplay. One mode makes the model think in first-person parenthetical asides—"(he noticed me… heart racing)"—while the other strips all inner monologue for cold scene analysis.

The interesting bit The trick exploits where DeepSeek was trained to inject instructions: the end of the first user message, not the system prompt. Because the full conversation history is re-fed each turn, a one-time append automatically governs every subsequent reply. The repo even provides a Python helper that concatenates the marker string and then forgets about it.

Key highlights

• Two modes: “role immersion” (actor-style inner monologue) and “pure analysis” (director-style planning)
• Works in DeepSeek’s expert mode, deepseek-v4-flash, and deepseek-v4-pro API; fast mode unsupported
• Probability-based, not deterministic—“if it fails, roll again”
• Includes a bonus “lottery” method to force thinking-chain openings by hijacking the <｜begin▁of▁thinking｜> token, explicitly noted as untrained and unreliable

Caveats

• Only Chinese documentation; no English translation in repo
• Effectiveness is probabilistic, not guaranteed
• The “lottery” token-hack is admitted to be speculative

Verdict Worth bookmarking if you’re building character-driven chatbots on DeepSeek-V4 and want inspectable, stylized reasoning chains. Skip it if you need deterministic behavior or don’t read Chinese.