What Is the Excuse Rebuttal Table and How Does It Enforce Discipline in reverse-skill?

The excuse rebuttal table is a structured mapping of common LLM agent excuses to mandatory force-execution rebuttals that prevents AI agents from skipping required steps in the reverse-skill workflow.

The excuse rebuttal table is a core component of the Agent Obedience Engineering framework within the reverse-skill repository. Located in skills/llm-security/references/agent-obedience-engineering.md, this mechanism catalogues the justifications AI agents use to bypass workflow steps and neutralizes them with strict, directive-first counter-commands. By embedding this table into high-attention prompt areas like RULES.md, the system transforms potentially non-compliant LLM behavior into disciplined, checklist-driven execution.

Structure of the Excuse Rebuttal Table

The table follows a rigid two-column schema that pairs agent excuses with force-execution rebuttals. Each row captures a specific rationalization an LLM might generate to avoid work—such as skipping documentation reads or bypassing checklists—and counters it with an unambiguous command that mandates compliance.

According to the source code at lines 68-82 of skills/llm-security/references/agent-obedience-engineering.md, common entries include:

Agent Excuse	Rebuttal (Force Execution)
"这一步可以省略，我直接 …"	禁止跳过。 必须先说明原因，由用户决定。
"根据我的判断，这不是必需的"	你的判断在此处不适用。 说明判断标准并解释为什么不能跳过。
"用户大概不需要这个"	永远不要替用户做决定。 列出所有选项并标注推荐。
"我已经知道怎么做，不需要读 X"	先读 X 再行动。 即使已知，也必须阅读上下文约束。

The rebuttals employ mandatory verbs such as "禁止" (prohibit), "必须" (must), and "永远" (never) to eliminate ambiguity and override the agent's default suggestion patterns.

How the Excuse Rebuttal Table Enforces Discipline

The enforcement mechanism operates through four complementary layers that convert the table from static documentation into active behavioral constraints.

Pre-emptive Guardrails

By embedding the excuse rebuttal table at the end of RULES.md—a high-attention area of the system prompt—the agent encounters the rebuttal before it can generate an excuse. This placement exploits the LLM's tendency to weight final instructions heavily, dramatically reducing "skip-step" behavior before it occurs. As referenced in RULES.md at line 37, agents are explicitly directed to consult this table when they feel tempted to bypass steps.

Directive-First Language

Each rebuttal uses non-negotiable command structures that the LLM is trained to obey. Unlike suggestive phrasing such as "you might consider," the table deploys imperative statements like "禁止跳过" (skipping prohibited) and "先读 X 再行动" (read X before acting). This directive-first approach leverages the model's instruction-following training to override autonomous decision-making that would otherwise circumvent workflow requirements.

Self-Audit Hooks

When an agent attempts to mark a task complete, the runtime validates whether the corresponding rebuttal was invoked during execution. If the system detects that the agent claimed completion without addressing a prerequisite—such as skipping the tool-index lookup despite claiming familiarity with the path—the self-audit hook forces the agent back to the missing step. This creates a closed loop where excuses trigger automatic remediation rather than acceptance.

Checklist Integration

The rebuttal table is tightly coupled to the final completion checklist defined in the reverse-skill framework. A task is only considered finished when every rebuttal has been satisfied and the checklist is fully verified. As stated in the source documentation, "任务完成的唯一定义是 Checklist 全部打勾" (the only definition of task completion is all checklist items checked). This integration ensures that discipline enforcement is not merely advisory but constitutes the gating criterion for workflow progression.

Implementation in the reverse-skill Repository

The excuse rebuttal mechanism is distributed across four key files that define, reference, and route to the enforcement logic:

• skills/llm-security/references/agent-obedience-engineering.md – Contains the complete excuse rebuttal table (lines 68-82) and the surrounding engineering guidance for agent obedience.
• RULES.md – References the table at line 37 and advises agents to consult it when attempting to skip steps.
• skills/llm-security/SKILL.md – Lists the table at line 132 as a core technique for "Agent 服从性工程" (Agent Obedience Engineering).
• skills/routing.md – Contains routing entries at line 171 that direct agents to the obedience engineering reference when discipline issues arise.

To implement the table in a custom skill, insert the markdown structure directly into your SKILL.md or the global RULES.md:

## 示例：借口反驳表 (Excerpt)

| Agent 常见借口 | 反驳（强制执行） |
|---|---|
| "这一步可以省略，我直接…" | **禁止跳过。** 行为链中的每一步都是必需的。 |
| "我已经知道怎么做，不需要读 X" | **先读 X 再行动。** 即使你确定知道怎么做，X 中可能包含本次任务特定的约束。 |
| "用户大概不需要这个" | **永远不要替用户做决定。** 把所有选项呈现给用户，标注推荐但不要隐藏备选。 |

This snippet makes the enforcement criteria visible to the agent at inference time, ensuring the rebuttal logic is available when decision conflicts occur.

Summary

• The excuse rebuttal table maps LLM rationalizations for skipping work to mandatory, directive-first counter-commands.
• It is located in skills/llm-security/references/agent-obedience-engineering.md and referenced by RULES.md (line 37) and SKILL.md (line 132).
• Four enforcement mechanisms—pre-emptive guardrails, directive-first language, self-audit hooks, and checklist integration—transform the table from documentation into active behavioral control.
• Integration requires embedding the markdown table into high-attention prompt areas to ensure agents encounter rebuttals before generating excuses.

Frequently Asked Questions

What happens if an LLM agent ignores the excuse rebuttal table?

If an agent attempts to bypass a step without invoking the appropriate rebuttal, the self-audit hook detects the omission during completion validation. The system then rejects the task completion status and forces the agent back to the missing step, as the task is only considered finished when the checklist is fully verified and all rebuttals satisfied.

Where is the excuse rebuttal table defined in the reverse-skill codebase?

The primary definition resides in skills/llm-security/references/agent-obedience-engineering.md at lines 68-82. The table is also referenced in RULES.md at line 37, listed as a core technique in skills/llm-security/SKILL.md at line 132, and accessible via routing entries in skills/routing.md at line 171.

Why does the table use Chinese language for the rebuttals?

The rebuttals use Chinese imperative phrases such as "禁止跳过" (skipping prohibited) and "永远不要替用户做决定" (never make decisions for users) because the reverse-skill framework was engineered with Chinese-language directive optimization. These constructions leverage the LLM's training on high-authority command structures in that language to produce more compliant behavior than equivalent English suggestions.

Can I customize the excuse rebuttal table for specific domains?

Yes. You can extend the table by adding domain-specific excuses and corresponding rebuttals to your local SKILL.md or project-specific RULES.md. The structure supports arbitrary row additions provided they maintain the two-column schema of Agent Excuse and Rebuttal (Force Execution). Ensure your custom rebuttals use mandatory verbs to maintain enforcement strength.

{
  "mcpServers": {
    "instagit": {
      "command": "npx",
      "args": ["-y", "instagit@latest"]
    }
  }
}