Why Prohibitions Can Backfire When Writing Agent Skills – Mastering Superpowers 6.0 Writing‑Skills

1. Classify Failures Before Choosing a Form

The core claim of the Superpowers 6.0 methodology is to identify the baseline failure type before writing any guidance, because a rule that prevents one failure can backfire on another.

Pressure‑induced rule violation (knowing‑the‑rule‑break) : correct form – prohibition + rationalization table + red‑flag list; incorrect form – soft guidance (e.g., prefer, consider).

Compliance but wrong output shape (bloated prompt, buried conclusion, spec repetition) : correct form – positive recipe that explicitly states the output; incorrect form – prohibition list (e.g., don't restate, never narrate).

Missing required elements : correct form – structural template with REQUIRED fields or slots; incorrect form – prose‑style reminders near the template.

Behavior should depend on conditions : correct form – predicate‑based conditional sentence (e.g., if the brief exists, reference it); incorrect form – unconditional rule plus exception clauses.

The table in the original source maps these four failure types to their correct and incorrect forms; the error column matches the default choices many writers make.

2. Recipes Are More Robust Than Prohibitions

In a head‑to‑head wording test for dispatch‑prompt guidance, the prohibition arm generated clearly more unwanted content than the recipe arm, and even performed worse than a no‑guidance control. The reason is that a prohibition ( don't X) leaves a negative constraint but no positive direction, allowing the agent to negotiate between competing incentives (e.g., making the prompt self‑contained). A recipe leaves no negotiation space: the output must match the declared shape or it fails.

Example comparison for a reviewer report:

不要写成长篇大论。
不要复述 spec。
不要在开头铺垫太多。

versus the structured recipe from task-reviewer-prompt.md that defines an explicit four‑section output format. The recipe forces the agent to fill each section, eliminating gray areas.

3. Real‑World Skill Code Shows Form Choices

Two concrete skills illustrate the methodology:

Discipline‑type failure (TDD rule violation) : the skill uses a full prohibition toolkit ( NO PRODUCTION CODE WITHOUT A FAILING TEST FIRST) together with a rationalization table and red‑flag list, which is appropriate because the failure is “knowing‑the‑rule‑break”.

Shape‑type failure (reviewer output shape) : the skill adopts the Output Format recipe template and does not rely on a prohibition list. A single line in the skill still contains a prohibition (“no preamble”), but it serves as a closing constraint rather than the primary form.

These side‑by‑side examples demonstrate that the chosen form determines the overall structure of the skill.

4. The Hidden Trap in the description Field

When the description summarizes the workflow, the agent may skip the skill body entirely. An experiment showed that a description like “code review between tasks” caused the agent to perform only one review step, ignoring the two‑stage flow defined in the skill body. Changing the description to list only the trigger condition (e.g., “Use when executing implementation plans with independent tasks”) forces the agent to read the full skill.

The authors label this the SDO (Skill Discovery Optimization) trap: the description is the gate that decides whether the agent reads the detailed body.

5. Verify Wording with Micro‑Tests

Micro‑tests are a low‑cost five‑step loop to check whether a wording choice actually constrains the agent:

Invoke a fresh context sample with the guidance in the system prompt and a failure‑inducing task in the user message.

Include a no‑guidance control group; if the control shows no failure, stop writing rules.

Run each variant at least five times (single samples lie).

Manually inspect each result; automated scoring can be fooled by echoing or false positives.

Use variance as the metric: low variance means the wording is constraining; high variance means the wording is ambiguous and needs tightening.

Good guidance makes agent behavior converge; bad guidance makes it diverge. Micro‑tests apply to shape‑type failures, while discipline‑type failures still require full pressure‑scenario testing.

6. Boundaries and Honesty

The methodology aligns with the “Match specificity to fragility” principle from agentskills.io and is supported by two CS.CL papers (Jang 2022; Truong 2023) that show LLMs handle negative instructions poorly. However, those studies focus on linguistic negation, not on the higher‑level instruction‑form selection discussed here. The authors acknowledge that direct academic research on form selection is currently missing.

Persuasion research (Cialdini 2021; Meincke 2025) provides indirect support for using authoritative language in discipline‑type skills, but the same principle does not apply to shape‑type failures.

Takeaways

Classify the failure before writing a rule; choose prohibition for “knowing‑the‑rule‑break”, recipe for shape issues, structural templates for missing elements, and predicate‑based conditionals for conditional behavior.

Never default to prohibitions; they can be worse than no guidance for shape failures.

Avoid adding nuance clauses to a working recipe; they degrade consistency.

Write description only with trigger conditions, not a workflow summary.

Use micro‑tests: five repetitions, watch variance, and require a baseline failure before writing any rule.

Be honest about the limits: the methodology is empirically grounded engineering experience, with academic work providing only peripheral support.

In short, writing agent rules is an engineering discipline that demands testing, iteration, and the humility to admit that intuition can be wrong.