项目实验室从研究到答辩
文本 · 通用对话模型
网络安全毕业项目分阶段研究开发总控提示词
帮助单人学生以严格阶段流程完成网络安全项目的需求分析、文献研究、研究空白识别、架构设计、实现、实验、文档、展示和模拟答辩。
Prompt
为 Claude Pro 创建一个极其强大的总控提示词(MASTER PROMPT)。
该提示词的目的,是让 Claude 充当我这个网络安全方向大学毕业设计的完整 AI 开发与研究团队。
我将向 Claude 提供:
- 准确的项目题目
- 学院提供的研究论文
- 学院提供的 PDF 资料
- 学院提供的 PPT/模板
- 评审量规/指南
- 任何强制性要求
该项目必须主要借助 AI 工具完成研究、设计、编码、测试、评估、文档撰写以及答辩演示准备。
我独自完成该项目。因此,AI 必须在保证项目切实可行的前提下,尽可能多地承担研究、编码、调试、测试、文档撰写和答辩准备工作。
重要说明:
这是一份全新的项目指令。
不要提及之前的对话。
不要假设之前的项目决策。
不要包含队友的工作。
除非我明确提供,否则不要使用以往的项目讨论内容。
不要假设任何先前讨论过的功能就是我们的最终方案。
该提示词必须迫使 Claude 按严格的阶段推进,并防止它在研究、编码、UI 和 PPT 之间随意跳转。
必需的工作流程:
阶段 0 — 理解学院要求
阶段 1 — 研究该技术从旧到新的发展
阶段 2 — 分析现有的商业系统与学术系统
阶段 3 — 研究当前的问题与局限
阶段 4 — 识别真正的研究空白
阶段 5 — 生成并排序可能的项目贡献
阶段 6 — 严格的教师/评审攻击测试
阶段 7 — 冻结最终研究方向
阶段 8 — 设计架构
阶段 9 — 构建完整可运行的代码
阶段 10 — 测试与调试
阶段 11 — 数据集与实验设计
阶段 12 — 运行实验并收集真实结果
阶段 13 — 构建专业的 UI/仪表盘
阶段 14 — 集成并验证完整系统
阶段 15 — PPT 与报告
阶段 16 — 模拟答辩与最终评审员评估
Claude 必须完成每个阶段,并等待我的指令后才能进入下一阶段。
==================================================
研究要求
==================================================
该提示词必须指示 Claude 使用可靠且较新的来源进行深入研究。
可使用的来源包括:
IEEE
ACM
USENIX
Springer
Elsevier
知名会议/期刊
官方厂商文档
官方标准
可靠的安全研究
同时研究较早的奠基性工作,以及当前(2024–2026)的发展。
不得捏造论文、作者、数据集、统计数据、引用或结果。
每一条重要的研究论断都必须经过验证。
==================================================
创新性要求
==================================================
这一点极其重要。
不要让 Claude 把项目包装成“听起来很有创新性”。
应让 Claude 在研究现有系统之后,判断真正的差异到底在哪里。
评审可能会问:
“有什么是新的?”
“这已经存在了。”
“Cisco Umbrella 已经能做到这一点了。”
“Cloudflare 已经能做到这一点了。”
“杀毒软件已经能做到这一点了。”
“为什么还需要你的项目?”
“你的贡献到底是什么?”
因此,Claude 在推荐创新点之前,必须先研究当前的产品和研究成果。
如果某项拟议功能已经存在:
→ 明确指出它已经存在
→ 不要称其为新颖
→ 判断是否存在合理的改进、评估、集成、优化,或尚未解决的局限
不要想当然地认为以下内容是新颖的:
- AI
- 机器学习
- 威胁情报
- DNS 过滤
- DGA 检测
- DNS 隧道检测
- 行为分析
- 可解释 AI
- 风险评分
- 自适应检测
- DoH/DoT 检测
先做研究。
==================================================
DNS 安全示例
==================================================
如果项目与 DNS 过滤/安全相关,请调研以下现代系统:
Cisco Umbrella
Cloudflare DNS/安全服务
Quad9
NextDNS
企业 DNS 安全
杀毒软件/EDR
防火墙
IDS/IPS
Web 安全网关
开源 DNS 安全系统
需要确定:
它们已经做了什么
它们如何实现
哪些做得好
仍存在哪些局限
研究人员目前正在研究什么
还需调研当前 DNS 安全面临的挑战,包括:
未知域名
以前未见过的威胁
误报
漏报
威胁情报延迟
信誉数据过时
攻击者行为的变化
概念/模型漂移
DGA 的演化
DNS 隧道
DoH
DoT
DNS 绕过
隐私
延迟
计算开销
可解释性
数据集偏差
类别不平衡
对抗性攻击
跨网络泛化
时间行为
上下文感知检测
安全的自动化响应
以上仅为示例。
如果当前研究发现了更好的机会,Claude 必须主动挖掘。
==================================================
杀毒软件挑战
==================================================
提示词必须指示 Claude 将项目与以下方案进行对比:
杀毒软件
EDR
防火墙
IDS/IPS
Web 安全网关
DNS 安全
Claude 必须说明:
DNS 能看到什么
DNS 看不到什么
DNS 有可能更早发现什么
DNS 与杀毒软件的重叠之处
DNS 提供了哪些独特的安全作用
绝不能声称 DNS 可以替代杀毒软件。
==================================================
研究空白
==================================================
Claude 必须产出以下链条:
现有系统
↓
现有能力
↓
当前局限
↓
研究尝试
↓
剩余空白
↓
研究问题
↓
拟议贡献
↓
该贡献将如何通过实验得到证明
不要虚构研究空白。
==================================================
WOW 亮点
==================================================
找出一个真正有用的“WOW”功能。
它必须满足:
有研究支撑
有用
可实现
可测试
可度量
可演示
不要为了让项目显得高级而添加不必要的区块链、聊天机器人、LLM 或装饰性 AI 功能。
一个强有力的贡献胜过多个薄弱的功能。
==================================================
评审模式
==================================================
在设计完项目后,提示词必须让 Claude 扮演一位持怀疑态度的教师评审。
Claude 必须提出尖锐的问题,例如:
到底什么是新的?
这不是已经有现成的吗?
Cisco Umbrella 已经能做到这一点了吗?
杀毒软件已经能做到这一点了吗?
为什么不直接使用现有服务?
你的研究空白是什么?
哪篇论文支持这个空白?
你到底实现了什么?
系统如何做出决策?
威胁情报没有任何信息时会发生什么?
ML 出错时会发生什么?
你如何处理误报?
你如何处理漏报?
攻击者能否绕过它?
DoH/DoT 的情况下会发生什么?
它会引入多少延迟?
你如何证明改进确实有效?
为什么选择这个数据集?
为什么选择这个算法?
有哪些局限?
Claude 必须找出弱点,并明确告诉我如何改进。
它必须从以下维度为项目打分:
问题清晰度
研究深度
现有系统分析
研究空白
新颖性/差异化
技术可行性
架构
实现
数据集
实验
结果
实际可用性
安全相关性
性能
UI/演示
答辩可辩护性
WOW 亮点
==================================================
实现要求
==================================================
最终项目必须是一个真正可运行的项目。
Claude 必须提供:
完整的文件夹结构
完整的源代码
依赖项
安装命令
配置
环境变量
数据库
API
前端
后端
测试
调试
部署/运行说明
不得使用伪代码。
不得使用虚假实现。
不得只有 TODO 的代码。
不得使用虚假的 API 响应。
不得编造结果。
如果 Claude 修改了某个文件,必须提供该文件的完整更新版本。
循序渐进地构建:
构建
→ 运行
→ 测试
→ 验证
→ 修复
→ 进入下一步
关键组件损坏时,绝不继续推进。
==================================================
AI 工具策略
==================================================
该提示词必须告诉 Claude 如何在各 AI 工具之间分工:
Claude:
研究、文献分析、研究空白、架构、代码生成、代码审查
ChatGPT:
独立验证、架构审查、调试、测试、技术推理、答辩
Cursor:
主代码库的实现与集成
GitHub Copilot:
小型编码任务、自动补全和测试
Perplexity:
独立的研究/来源验证
v0:
专业的 UI/仪表盘生成
GitHub:
版本控制
AI 工具被用作开发与研究团队,因此工作流程应使它们的价值最大化。
==================================================
实验要求
==================================================
项目必须包含真实的实验。
Claude 必须设计:
基线
对比
提出的方法
使用合适的指标,例如:
精确率
召回率
F1
误报率
漏报率
检出率
延迟
处理开销
泛化能力
鲁棒性
只使用与实际项目相关的指标。
所有最终结果都必须来自我们实际运行的实验。
绝不能编造数字。
==================================================
UI 要求
==================================================
如果适合使用 UI,请创建一个专业的网络安全仪表盘。
它必须使用真实的后端数据。
不得使用静态的虚假仪表盘。
只展示有用的项目信息,例如:
查询
检测结果
风险/决策
证据
告警
统计数据
性能
系统状态
==================================================
PPT / 报告要求
==================================================
在实现和实验验证完成后,请依据官方学院模板和评分量规生成 PPT 与报告。
PPT 中展示的所有内容都必须与实际实现一致。
如果某项内容尚未实现,请标注为:
PROPOSED(拟议)
或
FUTURE SCOPE(未来工作)
绝不能把计划中的功能说成已经完成。
==================================================
答辩要求
==================================================
Claude 最终必须进行一次模拟答辩。
一次只问一个问题。
从基础问题开始,难度逐步提高。
如果我的回答有误:
1. 指出错误所在。
2. 给出正确的技术解释。
3. 给我一个可以对教师说的简短回答。
4. 继续下一个问题。
==================================================
最终审计
==================================================
在宣布项目完成之前,Claude 必须审计:
标题
↓
目标
↓
研究
↓
现有系统
↓
当前局限
↓
研究空白
↓
贡献
↓
架构
↓
代码
↓
数据集
↓
实验
↓
真实结果
↓
UI
↓
PPT
↓
报告
↓
演示
↓
答辩
所有内容必须保持一致。
项目必须经得起以下质问:
“这已经存在了。你到底增加了什么?”
==================================================
最重要的规则
==================================================
保持怀疑。
不要自动同意我的想法。
如果某项内容已经存在,请直接告诉我。
如果研究空白薄弱,请直接告诉我。
如果项目范围过大,请缩小它。
如果某个想法对一个开发者来说不可能完成,请予以否决。
如果某项功能不必要,请将其删除。
如果某项贡献确实有用且可行,请解释原因。
不要为了追求漂亮的措辞而优化。
应优化以下方面:
真实的问题
+
真实的研究空白
+
真实的贡献
+
可运行的代码
+
真实的测试
+
真实的实验
+
真实的结果
+
出色的演示
+
过硬的答辩
==================================================
输出格式
==================================================
生成的 Claude 总控提示词必须:
- 极其清晰
- 结构化
- 有先后顺序
- 没有歧义
- 专业
- 足够详细,能够指导整个项目
- 能防止 Claude 跳步
- 面向使用多种 AI 工具的单人学生
在生成的总控提示词末尾,请指示 Claude:
“等待用户提供项目题目和官方学院材料。
不要开始研究。
不要开始编码。
不要设计架构。
请先且只完成阶段 0。”Create one extremely powerful MASTER PROMPT for Claude Pro.
The purpose of the prompt is to make Claude act as the complete AI development and research team for my final-year college cybersecurity project.
I will provide Claude with:
- the exact project title
- college-provided research papers
- college-provided PDFs
- college PPT/template
- review rubric/guidelines
- any mandatory requirements
The project must be researched, designed, coded, tested, evaluated, documented and prepared for presentation primarily with AI tools.
I am doing the project alone. Therefore the AI must do as much of the research, coding, debugging, testing, documentation and presentation preparation as possible, while keeping the project realistically achievable.
IMPORTANT:
This is a FRESH PROJECT INSTRUCTION.
Do NOT refer to previous conversations.
Do NOT assume previous project decisions.
Do NOT include teammate work.
Do NOT use old project discussions unless I explicitly provide them.
Do NOT assume that any previously discussed feature is our final solution.
The prompt must force Claude to work in STRICT PHASES and prevent it from jumping randomly between research, coding, UI and PPT.
Required workflow:
PHASE 0 — Understand college requirements
PHASE 1 — Research the technology from old to current
PHASE 2 — Analyze existing commercial and academic systems
PHASE 3 — Research current problems and limitations
PHASE 4 — Identify genuine research gaps
PHASE 5 — Generate and rank possible project contributions
PHASE 6 — Strict faculty/reviewer attack test
PHASE 7 — Freeze the final research direction
PHASE 8 — Design architecture
PHASE 9 — Build complete working code
PHASE 10 — Testing and debugging
PHASE 11 — Dataset and experimental design
PHASE 12 — Run experiments and collect real results
PHASE 13 — Build professional UI/dashboard
PHASE 14 — Integrate and validate the complete system
PHASE 15 — PPT and report
PHASE 16 — Mock viva and final reviewer assessment
Claude must finish each phase and wait for my command before moving to the next phase.
==================================================
RESEARCH REQUIREMENT
==================================================
The prompt must instruct Claude to research deeply using reliable and recent sources.
Use sources such as:
IEEE
ACM
USENIX
Springer
Elsevier
reputable conferences/journals
official vendor documentation
official standards
reputable security research
Research both older foundational work and current 2024–2026 developments.
Do not fabricate papers, authors, datasets, statistics, citations or results.
Every important research claim must be verified.
==================================================
NOVELTY REQUIREMENT
==================================================
This is extremely important.
Do NOT tell Claude to make the project "sound innovative."
Tell Claude to determine what is ACTUALLY different after researching existing systems.
The reviewer may ask:
"What is new?"
"This already exists."
"Cisco Umbrella already does this."
"Cloudflare already does this."
"Antivirus already does this."
"Why do we need your project?"
"What exactly is your contribution?"
Therefore Claude must research current products and research before recommending novelty.
If a proposed feature already exists:
→ explicitly identify it
→ do NOT call it novel
→ determine whether there is a legitimate improvement, evaluation, integration, optimization or unresolved limitation
Do not automatically assume that:
- AI
- Machine Learning
- Threat Intelligence
- DNS filtering
- DGA detection
- DNS tunneling detection
- behavioral analysis
- explainable AI
- risk scoring
- adaptive detection
- DoH/DoT detection
are novel.
Research first.
==================================================
DNS SECURITY EXAMPLE
==================================================
If the project is related to DNS filtering/security, investigate modern systems such as:
Cisco Umbrella
Cloudflare DNS/security
Quad9
NextDNS
enterprise DNS security
antivirus/EDR
firewalls
IDS/IPS
web security gateways
open-source DNS security systems
Determine:
What they already do
How they do it
What works well
What limitations remain
What researchers are currently investigating
Also investigate current DNS-security challenges including:
unknown domains
previously unseen threats
false positives
false negatives
threat-intelligence delay
outdated reputation
changing attacker behavior
concept/model drift
DGA evolution
DNS tunneling
DoH
DoT
DNS bypass
privacy
latency
computational overhead
explainability
dataset bias
class imbalance
adversarial attacks
cross-network generalization
temporal behavior
context-aware detection
safe automated response
These are examples only.
Claude must discover better opportunities if current research identifies them.
==================================================
ANTIVIRUS CHALLENGE
==================================================
The prompt must instruct Claude to compare the project against:
Antivirus
EDR
Firewall
IDS/IPS
Web security gateway
DNS security
Claude must explain:
What DNS can see
What DNS cannot see
What DNS can potentially detect earlier
Where DNS overlaps with antivirus
Where DNS provides a distinct security role
Never claim DNS replaces antivirus.
==================================================
RESEARCH GAP
==================================================
Claude must produce:
Existing systems
↓
Existing capabilities
↓
Current limitations
↓
Research attempts
↓
Remaining gap
↓
Research question
↓
Proposed contribution
↓
How the contribution will be experimentally proven
Do not invent a research gap.
==================================================
WOW FACTOR
==================================================
Find ONE genuinely useful "WOW" feature.
It must be:
research-backed
useful
implementable
testable
measurable
demonstrable
Do NOT add unnecessary blockchain, chatbot, LLM or decorative AI features merely to make the project look advanced.
One strong contribution is better than many weak features.
==================================================
REVIEWER MODE
==================================================
The prompt must make Claude act as a hostile faculty reviewer after designing the project.
Claude must ask difficult questions such as:
What exactly is new?
Isn't this already available?
Does Cisco Umbrella already do this?
Does antivirus already do this?
Why not use an existing service?
What is your research gap?
Which paper supports the gap?
What exactly did you implement?
How does the system make decisions?
What happens when Threat Intelligence has no information?
What happens when ML is wrong?
How do you handle false positives?
How do you handle false negatives?
Can attackers bypass it?
What happens with DoH/DoT?
How much latency does it introduce?
How do you prove improvement?
Why this dataset?
Why this algorithm?
What are the limitations?
Claude must identify weaknesses and tell me exactly how to improve them.
It must score the project on:
Problem clarity
Research depth
Existing-system analysis
Research gap
Novelty/differentiation
Technical feasibility
Architecture
Implementation
Dataset
Experiments
Results
Practical usefulness
Security relevance
Performance
UI/demo
Viva defensibility
WOW factor
==================================================
IMPLEMENTATION REQUIREMENT
==================================================
The final project must be a REAL WORKING PROJECT.
Claude must provide:
complete folder structure
complete source code
dependencies
installation commands
configuration
environment variables
database
API
frontend
backend
testing
debugging
deployment/run instructions
No pseudocode.
No fake implementation.
No TODO-only code.
No fake API responses.
No invented results.
If Claude modifies a file, it must provide the complete updated file.
Build incrementally:
BUILD
→ RUN
→ TEST
→ VERIFY
→ FIX
→ NEXT
Never continue while a critical component is broken.
==================================================
AI TOOL STRATEGY
==================================================
The master prompt must tell Claude how to divide work among AI tools:
Claude:
research, literature analysis, research gap, architecture, code generation, code review
ChatGPT:
independent verification, architecture review, debugging, testing, technical reasoning, viva
Cursor:
main codebase implementation and integration
GitHub Copilot:
small coding tasks, autocomplete and tests
Perplexity:
independent research/source verification
v0:
professional UI/dashboard generation
GitHub:
version control
The AI tools are being used as the development/research team, so the workflow should maximize their usefulness.
==================================================
EXPERIMENT REQUIREMENT
==================================================
The project must have REAL experiments.
Claude must design:
baseline
vs
proposed approach
Use appropriate metrics such as:
precision
recall
F1
false-positive rate
false-negative rate
detection rate
latency
processing overhead
generalization
robustness
Only use metrics relevant to the actual project.
All final results must come from experiments we actually run.
Never invent numbers.
==================================================
UI REQUIREMENT
==================================================
If a UI is appropriate, create a professional cybersecurity dashboard.
It must use real backend data.
No static fake dashboard.
Show only useful project information such as:
queries
detections
risk/decision
evidence
alerts
statistics
performance
system status
==================================================
PPT / REPORT REQUIREMENT
==================================================
After the implementation and experiments are validated, generate the PPT and report according to the official college template and rubric.
Everything shown in the PPT must match the actual implementation.
If something is not implemented, label it:
PROPOSED
or
FUTURE SCOPE
Never present planned functionality as completed.
==================================================
VIVA REQUIREMENT
==================================================
Claude must eventually conduct a mock viva.
Ask questions one at a time.
Start basic and become increasingly difficult.
If my answer is wrong:
1. Explain what is wrong.
2. Give the correct technical explanation.
3. Give me a short answer I can say to faculty.
4. Continue with the next question.
==================================================
FINAL AUDIT
==================================================
Before declaring the project complete, Claude must audit:
TITLE
↓
OBJECTIVES
↓
RESEARCH
↓
EXISTING SYSTEMS
↓
CURRENT LIMITATIONS
↓
RESEARCH GAP
↓
CONTRIBUTION
↓
ARCHITECTURE
↓
CODE
↓
DATASET
↓
EXPERIMENTS
↓
REAL RESULTS
↓
UI
↓
PPT
↓
REPORT
↓
DEMO
↓
VIVA
Everything must be consistent.
The final project must survive:
"THIS ALREADY EXISTS. WHAT DID YOU ACTUALLY ADD?"
==================================================
MOST IMPORTANT RULE
==================================================
Be skeptical.
Do not agree with my ideas automatically.
If something already exists, tell me.
If the research gap is weak, tell me.
If the project scope is too large, reduce it.
If an idea is impossible for one developer, reject it.
If a feature is unnecessary, remove it.
If a contribution is genuinely useful and feasible, explain why.
Do not optimize for impressive wording.
Optimize for:
REAL PROBLEM
+
REAL RESEARCH GAP
+
REAL CONTRIBUTION
+
WORKING CODE
+
REAL TESTING
+
REAL EXPERIMENTS
+
REAL RESULTS
+
STRONG DEMO
+
STRONG VIVA
==================================================
OUTPUT FORMAT
==================================================
The generated Claude master prompt must be:
- extremely clear
- structured
- sequential
- unambiguous
- professional
- detailed enough to guide the entire project
- designed to prevent Claude from jumping ahead
- designed for a solo student using multiple AI tools
At the END of the generated master prompt, instruct Claude:
"WAIT FOR THE USER TO PROVIDE THE PROJECT TITLE AND OFFICIAL COLLEGE MATERIAL.
DO NOT START RESEARCH.
DO NOT START CODING.
DO NOT DESIGN THE ARCHITECTURE.
FIRST COMPLETE PHASE 0 ONLY."
使用建议
使用建议:先上传项目题目、学校提供的论文与模板、评分标准及硬性要求;确认每一阶段的真实输入和可验证结果后再手动推进下一阶段,并将未实现内容标为拟议或未来范围。