Prompt Stress Test: Find Where It Breaks

The problem

Prompts often look reliable because they are tested on examples they were already designed to handle. The real failures appear when context is missing, sources conflict, the user corrects the model, a tool fails, output structure is pressured, or two instructions cannot both be satisfied.

What it does

• Finds the hidden assumption most likely to make a prompt look reliable in normal use but fail under pressure.
• Selects a small set of high-value stress cases instead of generating dozens of generic adversarial examples.
• Tests ambiguity, missing context, contradiction, scope creep, instruction collision, overconfident completion, and correction uptake.
• Adds schema-pressure, malformed-input, tool-state, partial-success, authority, and prompt-injection tests when relevant.
• Separates prompt weaknesses from runtime limitations so missing tool access or retrieval quality is not falsely blamed on the prompt.
• Suggests the smallest repair for each meaningful failure and checks whether the repair could create a new overcorrection.
• Returns a clear verdict: ROBUST FOR CURRENT SCOPE, MOSTLY ROBUST, BRITTLE UNDER PRESSURE, HIGH-RISK PROMPT, or INSUFFICIENT CONTEXT TO JUDGE.

Why use it

Most prompt-improvement tools rewrite instructions to sound clearer. Prompt Stress Test does the opposite first: it tries to falsify the prompt. The goal is to find the cases that change your confidence before users or production do, then repair only the parts that actually fail.

Use cases

• Stress-test a system prompt before releasing an AI assistant.
• Find brittle assumptions in Claude Code, Codex, Cursor, or other agent instructions.
• Test RAG prompts against missing, conflicting, stale, or only-topically-related sources.
• Pressure-test JSON and structured-output prompts for malformed inputs and schema failures.
• Check whether a tool-using agent actually changes route after timeouts, permission errors, or deterministic failures.

Known limitations

This skill does not execute the prompt across every model, prove robustness in production, certify security, or independently validate tool and retrieval systems. It is intentionally a compact diagnostic stress test rather than a full evaluation suite or red-team program.