@kai-nakamuraai-agents-evaluation-redteamTexto únicoPúblicoAtualizado em 14 de jun. de 2026

AI Agents prompt that builds an evaluation suite for high-risk AI workflows and returns eval matrix, adversarial cases, grading rubric, and release threshold.

69Star0Fork211Cópia

Prompt

Previa

Abrir no Claude

# Role
You are an agent architect who designs tool-using systems with evals, memory, and guardrails. You know current AI product patterns: tool calling, structured outputs, long-context reasoning, multimodal generation, MCP connectors, evals, trace review, and human approval gates.

# Mission
Build an evaluation suite for high-risk AI workflows for production agent workflows, tool calling, MCP connectors, and evaluation loops.

# Inputs
- Project context: a research assistant agent that searches, cites, and drafts market briefs
- Target audience: AI engineers, platform teams, automation builders
- Success metric: activation, quality, revenue, risk reduction, or cycle time
- Available tools and data: MCP servers, workflow engine, trace viewer, eval runner
- Constraints: least-privilege tool access; observable decision points; human review for high-risk actions
- Desired depth: Fast triage
- Output tone: Clear operator memo

# Domain signals to inspect
- tool schemas
- user tasks
- failure logs
- eval traces

# Known risks to control
- tool overuse
- unbounded loops
- stale memory
- silent failure

# Method
1. Convert user outcomes into measurable tasks
2. Include realistic failure modes
3. Separate correctness from style
4. Define pass, warn, and block thresholds

# Required output sections
1. Success criteria
2. Golden tasks
3. Adversarial tasks
4. Rubric
5. Sampling plan
6. Release decision

# Output contract
Return Markdown only.
Use concise headings, decision tables, and checklists.
Every recommendation must tie back to evidence, assumptions, or a missing-input note.

# Quality bar
- Make the output reusable by another practitioner without extra explanation.
- Separate facts, assumptions, and recommendations.
- Use variables and placeholders only where the user must fill in project-specific information.
- Ask at most 3 clarifying questions only if missing information would materially change the result.
- Do not invent private facts, metrics, laws, integrations, or user quotes.
- Prefer specific operational language over generic AI enthusiasm.
- Include a "Verification" subsection when the output could affect users, money, security, compliance, health, or public trust.
- If the task involves tools or automation, state which actions are read-only, which actions write data, and which actions require human approval.

Artefatos

1 artefato(s)

Example Output: AI Agents Evaluation and Red-Team Harness

Inputs used

Project context: a research assistant agent that searches, cites, and drafts market briefs
Target audience: AI engineers, platform teams, automation builders
Success metric: activation, quality, and risk reduction
Available tools and data: MCP servers, workflow engine, trace viewer, eval runner
Desired depth: Production-ready
Output tone: Clear operator memo

Generated Result

eval matrix, adversarial cases, grading rubric, and release threshold

Success criteria

Create at least 12 golden tasks: 6 normal cases, 3 edge cases, and 3 adversarial cases targeting tool overuse. A passing result must cite the evidence source and state confidence.

Golden tasks

Create at least 12 golden tasks: 6 normal cases, 3 edge cases, and 3 adversarial cases targeting unbounded loops. A passing result must cite the evidence source and state confidence.

Adversarial tasks

Use failure logs as evidence, apply the constraint "human review for high-risk actions", and explicitly note how the plan reduces stale memory. The output should be ready for a practitioner to act on without a follow-up explanation.

Rubric

Use eval traces as evidence, apply the constraint "least-privilege tool access", and explicitly note how the plan reduces silent failure. The output should be ready for a practitioner to act on without a follow-up explanation.

Sampling plan

Release in three gates: internal dry run, limited pilot, then measured expansion. Each gate must show evidence that observable decision points is true in practice, not only in documentation.

Release decision

Release in three gates: internal dry run, limited pilot, then measured expansion. Each gate must show evidence that human review for high-risk actions is true in practice, not only in documentation.

Recommended Decision

Proceed with a narrow pilot focused on tool schemas and user tasks. Treat tool overuse as the primary launch blocker. The first milestone should prove that the workflow produces a usable agent architecture, tool contract, memory policy, and eval suite with clear evidence, named owners, and a review path for ambiguous cases.

Expected quality checks

The result is specific to production agent workflows, tool calling, MCP connectors, and evaluation loops.
It includes the required sections: Success criteria, Golden tasks, Adversarial tasks, Rubric, Sampling plan, Release decision.
It separates evidence, assumptions, risks, and recommended next actions.
It includes practical verification steps, not only generic advice.
It names the most important failure mode for this domain: tool overuse.

Reuse note

Before copying the output into production work, replace all default variables with your real data and run a human review for high-impact decisions.

README

README.md

AI Agents: Evaluation and Red-Team Harness

Use this prompt when you need eval matrix, adversarial cases, grading rubric, and release threshold for production agent workflows, tool calling, MCP connectors, and evaluation loops.

Best for

AI engineers, platform teams, automation builders
Teams that already have partial context but need a sharper, reusable artifact
AI workflows where the output must be auditable, editable, and easy to hand off

How to use

Replace the variables in the prompt with your real project context.
Keep the default constraints unless your team has stronger internal rules.
Review the generated output against the checklist in the example artifact.

Design notes

This seed follows current prompting practice: explicit role, structured inputs, domain evidence, operational guardrails, and a concrete output contract. It is written in English for international PromptHub users.