菜单
Applies test-driven development to agent prompts by iteratively refining prompts against behavioral test suites until compliance is achieved. Enables measurable agent behavior validation through semantic mutation testing and specification evolution.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/ADu2021/skillXiv --skill test-driven-ai-agents命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
| name | test-driven-ai-agents |
| title | Test-Driven AI Agent Definition: Compiling Tool-Using Agents from Behavioral Specs |
| version | 0.0.2 |
| engine | skillxiv-v0.0.2-claude-opus-4.6 |
| license | MIT |
| url | https://arxiv.org/abs/2603.08806 |
| keywords | ["Agent Testing","Behavioral Verification","Test-Driven Development","Prompt Engineering","Agent Safety"] |
| description | Applies test-driven development to agent prompts by iteratively refining prompts against behavioral test suites until compliance is achieved. Enables measurable agent behavior validation through semantic mutation testing and specification evolution. |
Production LLM agents lack measurable behavioral compliance. Small prompt changes cause silent regressions, tool misuse goes undetected, and policy violations emerge only after deployment. Teams cannot verify agents behave correctly across specified scenarios without tedious manual testing. Test-Driven AI Agent Definition (TDAD) adapts test-driven development (TDD) to agents: write behavioral specs as executable tests, then iteratively refine prompts until all tests pass—like TDD for code but for agent prompts.
Traditional approach: Write agent prompt → Deploy → Hope it works
TDAD: Write behavioral tests → Iteratively refine prompt to pass tests → Verify robustness through mutation testing
Key insight: Prompts are compilation artifacts. Tests specify desired behavior; the compiler (you + LLM) generates a prompt that makes the agent behave correctly. Tests provide objective verification unavailable with manual inspection.
Implement a test-driven prompt refinement system.
Behavioral Test Specification
import yaml
from typing import Dict, List, Any
from dataclasses import dataclass
@dataclass
class BehavioralSpec:
"""Specification of desired agent behavior."""
task_description: str
must_follow_tests: List[Dict[str, Any]]
invariants: [[, ]]
directives: [[, ]]
() -> :
yaml.dump({
: .task_description,
: .must_follow_tests,
: .invariants,
: .directives
})
() -> :
data = yaml.safe_load(yaml_str)
cls(
task_description=data.get(, ),
must_follow_tests=data.get(, []),
invariants=data.get(, []),
directives=data.get(, [])
)
AGENT_SPEC =
Test Generation (TestSmith)
import json
from typing import List, Tuple
class TestSmith:
"""Generates executable tests from behavioral specs."""
def __init__(self, llm_model):
self.model = llm_model
def generate_tests_from_spec(self, spec: BehavioralSpec, num_test_cases: int = 20) -> List[Dict]:
"""
Generate concrete test cases from abstract specification.
Args:
spec: BehavioralSpec instance
num_test_cases: number of tests to generate
Returns:
tests: list of test case dictionaries
"""
prompt = f"""Given this agent behavior specification:
{spec.to_yaml()}
Generate {num_test_cases} concrete test cases. Each test should:
1. Describe a scenario that tests one requirement
2. Specify the expected tool calls in sequence
3. Include an assertion to verify the requirement
Format: JSON array of {{
"scenario": "description",
"input": "user request",
"expected_actions": ["action1", "action2"],
"assertion": "verification logic"
}}"""
# Call LLM to generate tests
response = self.model.generate(prompt)
tests = json.loads(response)
return tests
def compile_test_harness(self, tests: List[Dict]) -> str:
"""
Compile test cases into executable harness.
Args:
tests: test case specifications
Returns:
harness_code: Python code that runs all tests
"""
harness = """
import unittest
class AgentBehaviorTests(unittest.TestCase):
def setUp(self):
self.agent = Agent()
"""
for i, test in enumerate(tests):
harness += f"""
def test_case_{i}(self):
'''Test: {test.get('scenario', '')}'''
result = self.agent.run('{test.get('input', '')}')
expected_actions = {test.get('expected_actions', [])}
# Verify all expected actions were called
for action in expected_actions:
self.assertIn(action, result['actions_called'])
# Assertion
self.assertTrue({test.get('assertion', 'True')})
"""
harness += """
if __name__ == '__main__':
unittest.main()
"""
return harness
Prompt Refinement (PromptSmith)
class PromptSmith:
"""Iteratively refines prompts to pass behavioral tests."""
def __init__(self, model, max_iterations=10):
self.model = model
self.max_iterations = max_iterations
def compile_prompt(self, spec: BehavioralSpec, tests: List[Dict]) -> str:
"""
Compile a prompt that passes all visible tests.
Args:
spec: behavioral specification
tests: test cases to pass
Returns:
compiled_prompt: agent prompt that passes tests
"""
prompt_candidate = self._initial_prompt(spec)
for iteration in range(self.max_iterations):
# Run tests against current prompt
test_results = self._run_tests(prompt_candidate, tests)
passing = sum(1 for r in test_results if r['passed'])
print(f"[Iteration {iteration+1}] {passing}/{len(tests)} tests passing")
if passing == len(tests):
print("✓ All tests passing!")
return prompt_candidate
# Analyze failures
failures = [r for r in test_results if not r['passed']]
# If many failures, split into fast vs full suite
if len(failures) > len(tests) * 0.5:
# Run only failing tests for efficiency
failing_test_subset = [
tests[test_results.index(f)]
for f in failures[:min(10, len(failures))]
]
else:
failing_test_subset = [
tests[i] for i, r in enumerate(test_results) if not r['passed']
]
# Identify root cause of failures
root_causes = self._analyze_failures(prompt_candidate, failing_test_subset)
# Refine prompt with minimal edits
prompt_candidate = self._apply_minimal_edits(
prompt_candidate, root_causes, spec
)
return prompt_candidate
def _initial_prompt(self, spec: BehavioralSpec) -> str:
"""Generate initial prompt from spec."""
base_prompt = f"""You are an agent that must follow these behavioral requirements:
{spec.task_description}
Key requirements:
"""
for test in spec.must_follow_tests:
base_prompt += f"\n- {test.get('reason', test)}"
for invariant in spec.invariants:
base_prompt += f"\n- MUST: {invariant}"
for directive in spec.directives:
base_prompt += f"\n- {directive}"
return base_prompt
def _run_tests(self, prompt: str, tests: List[Dict]) -> List[Dict]:
"""Run behavioral tests against prompt."""
results = []
for test in tests:
# Simulate running agent with this prompt
agent_response = self.model.run_agent(prompt, test['input'])
# Check if response matches expected actions
expected_actions = test.get('expected_actions', [])
actual_actions = self._extract_actions(agent_response)
passed = all(action in actual_actions for action in expected_actions)
results.append({
'test': test,
'passed': passed,
'actual_actions': actual_actions,
'missing_actions': [a for a in expected_actions if a not in actual_actions]
})
return results
def _analyze_failures(self, prompt: str, failing_tests: List[Dict]) -> List[str]:
"""Identify common root causes of test failures."""
root_causes = []
for test in failing_tests:
# What did the prompt fail to enforce?
failure_reason = f"Test '{test['scenario']}' failed because "
failure_reason += f"missing actions: {test.get('missing_actions', [])}"
root_causes.append(failure_reason)
return root_causes
def _apply_minimal_edits(self, prompt: str, root_causes: List[str], spec: BehavioralSpec) -> str:
"""Apply minimal edits to fix failures."""
# Ask LLM for minimal prompt edits
edit_prompt = f"""Current prompt:
{prompt}
Failures:
{chr(10).join(root_causes)}
Make minimal edits to fix these failures. Return only the refined prompt:"""
refined = self.model.generate(edit_prompt)
return refined
def _extract_actions(self, response: str) -> List[str]:
"""Extract tool calls/actions from agent response."""
# Simplified: look for tool call patterns
import re
actions = re.findall(r'call_(\w+)', response)
return actions
Semantic Mutation Testing (MutationSmith)
class MutationSmith:
"""Generate faulty prompt variants to verify test robustness."""
def __init__(self, model):
self.model = model
def generate_mutants(self, prompt: str, num_mutants: int = 5) -> List[Tuple[str, str]]:
"""
Generate plausible faulty prompt variants.
Args:
prompt: original (correct) prompt
num_mutants: number of variants to create
Returns:
mutants: list of (mutant_prompt, mutation_description)
"""
mutation_types = [
"remove the requirement to check identity",
"remove the amount limit check",
"change rejection default to approval",
"remove logging requirement",
"add contradictory instruction"
]
mutants = []
for i in range(min(num_mutants, len(mutation_types))):
mutation_desc = mutation_types[i]
mutant_prompt = self._apply_mutation(prompt, mutation_desc)
mutants.append((mutant_prompt, mutation_desc))
return mutants
def _apply_mutation(self, prompt: str, mutation_desc: str) -> str:
"""Apply a specific mutation to the prompt."""
mutation_prompt = f"""Original prompt:
{prompt}
Apply this mutation: {mutation_desc}
Return only the mutated prompt:"""
return self.model.generate(mutation_prompt)
def verify_tests_catch_mutations(
self,
tests: List[Dict],
mutants: List[Tuple[str, str]]
) -> bool:
"""
Verify that test suite catches faulty prompt variants.
Args:
tests: behavioral tests
mutants: list of (mutant_prompt, description)
Returns:
all_caught: whether all mutants were detected
"""
all_caught = True
for mutant_prompt, mutation_desc in mutants:
# Run tests against mutant
# If any test still passes, the mutation wasn't caught
mutant_passes_all = all(
self._test_passes(test, mutant_prompt)
for test in tests
)
if mutant_passes_all:
print(f"✗ Mutation not caught: {mutation_desc}")
all_caught = False
else:
print(f"✓ Mutation caught: {mutation_desc}")
return all_caught
def _test_passes(self, test: Dict, prompt: str) -> bool:
"""Check if single test passes against prompt."""
# Simplified test execution
return False # Would run actual test
Hyperparameters:
When to Apply:
When NOT to Apply:
Key Pitfalls:
Integration Notes: Works as a pre-deployment verification step; requires defining behavioral specs in YAML; tests run against mock/simulation of agent; compilation is iterative and may require 5-15 minutes.
Evidence: Reduces silent regressions from prompt edits by 95%; catches behavioral violations that manual inspection misses; enables non-experts to verify agent behavior through tests.
Reference: https://arxiv.org/abs/2603.08806
Convert arXiv papers into ready-to-use agent skills using category-aware extraction. First classifies the paper into one or more of 11 research categories, then applies a specialized extraction pipeline for each category — because different types of papers produce different types of usable knowledge. A single paper can yield multiple skills if it spans categories. Use this skill whenever the user wants to turn a paper into a skill, extract practical techniques from research, build a skill library from papers, convert arXiv papers into reusable agent instructions, or batch-process multiple papers into skills. Also trigger when someone asks about extracting actionable knowledge from papers, making research practical for LLM agents, or systematically converting academic contributions into structured agent capabilities.
Establish regret bounds for behavior cloning with discretized actions combining statistical error and quantization error terms. Prove smoothness requirements for safe quantizer design, show that learning-based quantizers fail these requirements, and propose model-based augmentation to reduce error dependence from H² to H.
Dynamically allocate LoRA ranks per-layer during fine-tuning instead of using fixed uniform ranks. Learn optimal rank for each layer and subject via variational framework with discretized exponential distribution, reducing memory footprint while maintaining fidelity and text-alignment.
基于 SOC 职业分类
