Handbook for Operating Reasoning Models: A Systematic Methodology from "Chatting" to "Controlling"

"LLM Operators Handbook Vol.1" is a practical guide addressing pain points in interacting with reasoning-based large language models. Its core concept is to shift from "chatting with AI" to "controlling a reasoning engine". The handbook adopts a pragmatic structure: from theoretical mechanisms (how models work) → system prompt design (rule setting before dialogue) → operational practice (prompt writing and interaction optimization) → fault recovery (intervention for out-of-control scenarios) → quick reference tools (quick-start templates), helping users efficiently master reasoning models.

With the popularization of reasoning models like OpenAI's o-series and DeepSeek-R1, the traditional "chatting" mode (casual questioning) easily leads to negative outcomes: endless reasoning loops, perfectionism-induced delayed outputs, and resource waste. This handbook addresses this pain point by requiring users to deeply understand model mechanisms, master systematic prompt engineering, and intervene manually when necessary, achieving a paradigm shift from passive questioning to active control.

Inside reasoning models, there is a dynamic game between the generator (producing candidate answers) and the critic (evaluating and questioning). When the critic is overactivated, it triggers a "perfectionism loop": the model repeatedly self-doubts, rejects conclusions, and fails to output results. This loop is often triggered by subjective superlative adjectives in prompts (such as "best" or "perfect"). Understanding this mechanism is the core foundation for subsequent strategies.

The handbook introduces the concept of "lexical engineering", which influences model weights through four categories of words:

1. Binding/Deterministic Words: Establish constraint contracts to prompt the model to stop self-doubt;
2. Sufficient/Pragmatic Words: Switch to a "pursue usability" mode and suppress the critic;
3. Forbidden/Prohibited Words: Set boundaries to prevent recursive reasoning;
4. Precedence/Override Words: Establish a hierarchy of needs to avoid decision paralysis.

System prompts are the "game rules" before dialogue, with core elements including:

1. Expert Peer Persona: Require the model to output technical details as a professional peer to reduce redundancy;
2. Anti-Loop Framework: Enforce linear processing and prohibit recursive self-review;
3. Cognitive Closure Mechanism: Require the model to attach a completion statement (e.g., "No further optimization needed") to clarify the termination signal. A well-designed system prompt can multiply interaction efficiency several times over.

Single prompts need to avoid activating the critic:

• Avoid Superlative Adjective Traps: Replace subjective words with objective constraints (e.g., "cost below X" instead of "best solution");
• Clearly Define Trade-off Relationships: State goal priorities (e.g., quality > cost > speed);
• Distinguish Between Constraints and Goals: Clarify hard constraints (non-violable) and subjective goals (compromisable). For complex tasks, modular step-by-step processing (compilation unit method) is recommended: interface definition → logic implementation → optimization, combined with few-shot scaffolding to lock the output style.

When reasoning gets out of control, you need to:

1. Identify Spiral Signals: Such as repeated "Actually...wait", long outputs without conclusions, or questioning the task definition;
2. Hard Reset Command: Send "Stop current reasoning and re-answer based on simplified requirements" to forcefully break the loop;
3. Constraint Injection: Directly make choices for the model (e.g., "Choose plan A and implement it immediately") to push the task forward. Fault recovery is a normal operation step, not a remedial measure.

Value of the handbook: stable output quality, controllable reasoning process, and efficient resource utilization. Quick reference tools include:

• Superlative Adjective Translation Table: Convert subjective words into objective constraints;
• Emergency Intervention Command List: Standardized reset/injection templates;
• Anti-Loop System Prompt Template: Reusable framework. The handbook promotes the evolution of users from passive questioners to active operators. Its open-source nature will continuously absorb community practices, making it a core resource for AI users to upgrade their skills.