Agent Prompt Stress-Tester

What this stress-tester actually checks

A good agent prompt is not about clever wording - it is about structure. Agents that misbehave almost always have a structural gap: no clear role, no bounded scope, no definition of done, no guardrails for ambiguity. This tool scans your pasted prompt for the signals that those structural elements are present, scores each one, and tells you exactly what is missing and how to fix it. It runs entirely in your browser: nothing is sent to a server, no model is called, and your prompt never leaves the page.

Because it is a heuristic, a high score means the prompt is well-formed, not that it is guaranteed to work. It is a linter, not a judge. But in practice, prompts that pass all seven checks fail far less often than prompts that skip them, because the failures they catch are the ones that produce vague, drifting, or dangerous agent behavior.

The seven things a solid agent prompt needs

• Clear role - the agent knows who it is. 'You are a security-focused code reviewer' anchors every decision it makes.
• Scoped responsibility - what it should do and, just as important, what it should not. Unbounded agents wander into unrelated work.
• Explicit process - a short ordered procedure so behavior is consistent from run to run.
• Success criteria - a definition of done, so the agent stops at the right point instead of arbitrarily.
• Output format - the shape of the reply (a diff, a bullet list, JSON), especially when something downstream depends on it.
• Guardrails - what to do when unclear ('state your assumption, do not guess') and what it must never do (force push, delete, spend).
• Examples - at least one concrete sample that resolves the ambiguity instructions alone leave behind.

Why scope and guardrails carry the most weight

The two failures that cause the most real damage are unbounded scope and missing guardrails. An agent with no scope will happily 'help' by editing files it was never asked to touch. An agent with no guardrails will guess when it should have asked, or run a destructive command because nothing told it not to. These are not style problems - they are the difference between an agent you can trust to run and one you have to babysit. That is why this tool weights them heavily and why fixing them first gives the biggest reliability gain.

How to read your score

1. 80-100 (Strong) - the prompt has the structural bones of a reliable agent. Focus your remaining effort on the content of each section, not its presence.
2. 55-79 (Decent) - workable but missing one or two important elements. Fix the gaps flagged in red first; they are usually scope, success criteria, or guardrails.
3. Below 55 (Weak) - the prompt is likely to produce inconsistent or drifting behavior. Rebuild it around the seven elements before relying on the agent.
4. Remember the score measures structure, not correctness. A well-structured prompt with wrong instructions still scores high - you still have to read it.

Turning gaps into fixes

Each failed check comes with a specific, copyable fix - not just 'add guardrails' but the actual sentence pattern to add. Work top to bottom: role first, then scope, then process, then the done-criteria and output format, then guardrails and examples. You do not need a long prompt to pass; you need a complete one. Many of the best agent prompts are under 200 words but hit all seven elements crisply. Longer is not better, because every line is loaded on each run and costs tokens.

Where this fits in building agents

Use this after you draft a subagent with the Subagent Generator and before you save it to .claude/agents/. Paste the body, read the gap report, patch the holes, and only then commit the file. It is the cheap quality gate that catches the mistakes you would otherwise only discover when the agent misbehaves mid-task. Pair it with tool scoping - a tight prompt plus a minimal tool set is what makes a subagent both safe and reliable.

What it deliberately does not do

This tool does not run your prompt, call a model, or judge whether your instructions are correct for your goal. It cannot tell you that your reviewer should also check for SQL injection, or that your test-writer targets the wrong framework - that is on you. It checks that the prompt is well-formed, which is necessary but not sufficient. Treat a strong score as permission to ship the structure, then still read the content with your own eyes.