Huang Yihe's full accumulation on prompt engineering + context engineering from 2024-03 to 2026-05. Covers 186 posts in the prompt-engineering topic bucket + cross-bucket cross-referenced materials. Focuses on "how non-technical individuals can use Prompt to leverage LLM output quality".
When to open:
- You want to learn "Prompt engineering" but don't know if it's worth the time
- You want to build product-level Prompt orchestration (not just template reuse)
- You want to understand the Prompt → Context → Skill evolution path
- You are stuck on selection questions like "Thinking models are strong enough, do I still need Sequential Thinking?"
- You do Vibe Coding / text-to-image / structured generation and want reusable Prompt patterns
0. Core stance: Prompt engineering does not need to be "learned" (2025-05 posts 1247, 1223)
The counter-consensus view Lao Huang repeatedly emphasizes:
"For ordinary users,there is absolutely no need to deliberately learn prompts, let alone buy courses. If you learn logic well and master some expression and communication methodologies, that is enough to organize language, drive model reasoning, and get expected results."
"What is the essence of prompts?Logic, methodology. The most common prompts we use tell the AI:who it is, what the background is, what to do, how to do it, and the success image. Is this long string some special course? Isn't it just breaking down a task, writing a clear Brief? If a new person joins the company and I need to onboard them, that's exactly how I teach them too."
Why this stance:
- The essence of prompts = task breakdown + Brief writing(the core skill of Lao Huang, a marketing veteran)
- Most "prompt courses" on the market are just packaging → you can't extrapolate after learning them
- The real way to improve:Learn logic / the Pyramid Principle / communication methodologies
Anti-patterns:
- Treating "I want to learn Prompt engineering" as the goal → 90% of time will be wasted
- Collecting hundreds of Prompt templates → most scenarios can't use them, they become useless as soon as the scenario changes
- Learning "5 secrets for AI to write Xiaohongshu posts" from bloggers → it has nothing to do with the blogger, breaking down your own requirements clearly is the key
When to revisit: Come back and read this when you want to buy a course. Use that money to subscribe to Claude / Gemini Pro for a year, then figure it out yourself.
1. The underlying mindset for communicating with models (2024-12 post 661)
Core metaphor:
"Treat the model as a new hire who writes well but doesn't know how to write Xiaohongshu posts. Just think: how would you teach it? Then experiment and revise repeatedly."
Key operations:
- Treat the other party as a smart new hire who lacks background, not a search engine or an encyclopedia
- Tell it what you want(goal + success image), not just keywords
- Provide necessary background(industry, style, boundaries, what to avoid)
- Iterate repeatedly: if it doesn't work once, revise the Prompt, don't scold the model for being stupid
When to use:
- Whenever you get confused writing a Prompt, ask "How would I explain this to a new hire who doesn't know anything about this?"
- When you feel the model "doesn't understand", ask yourself "Did I not explain clearly, or is the model really incapable?" (99% of the time it's the former)
2. Classic Prompt template: 5 elements of task breakdown (2025-05 post 1223)
The "five elements of classic prompts" given by Lao Huang:
| Element | Content |
|---|---|
| Who it is | Role setting (you are a financial analyst / translation expert /...) |
| What background | Context (what skills you have, what resources you have, what domains you understand) |
| What to do | Goal (what task you need to complete) |
| How to do it | Implementation path (how many steps, what method to use, what the constraints are) |
| Success image | Reference cases (what similar good works look like, or few-shot examples) |
Why this set covers 90% of scenarios: It is exactly the SOP for "how to onboard a new hire". It is universal at the logic level; professional differences only exist at the scenario level.
3. Best practices for translation/summarization Prompts (2024-12 post 693)
The "web search enhancement" Prompt that Lao Huang commonly uses:
Always reply in Chinese. Prefers English sources for online searches,
with Chinese sources as a last option. Always search the web for English
pages, even if questions are in Chinese. Translate questions to English
before searching.
Why: The quality of Chinese and English source information is worlds apart. For the same question, the answer density and accuracy from English sources are often 3-5 times that of Chinese sources. This Prompt makes "search in English + reply in Chinese" the default behavior.
When to use:
- Save it in Prompt House or Claude's Custom Instructions / Gemini's System Prompt for long-term effect
- Pair it with web-connected tools like Gemini / Perplexity, and all cited sources from searches will prefer English
4. iPhone Shortcuts + Prompt (2024-05 highlights from post 263)
3 iCloud Shortcuts shared by Lao Huang(using GPT-4o + Markdown output):
- Q&A GPT: voice input → speech-to-text + Prompt → GPT-4o → Notes
- Inspiration GPT: same as above
- Summary GPT: open WeChat article in Chrome → share to Shortcut → GPT-4o extracts key points → Notes
Key Prompt design:
- Force Markdown format output(convenient for later editing in Obsidian / Ulysses)
- Save to Notes (fast cross-device sync, can continue editing immediately on Mac)
Why this design works:
- "AI + system shortcuts" has deeper integration than "AI Apps"
- Markdown is the "universal language" of the AI era; forcing MD output by default in any Prompt makes downstream processing 10 times easier
5. Thinking-Claude: performing thinking is also valuable (2024-11 highlights from posts 614, 616)
Event: The browser extension Thinking-Claude adds "slow thinking" prompts to Claude, and the effect was so good that Lao Huang called it an "epic enhancement" on the first day.
The next day Lao Huang retracted his judgment:
- It is not real thinking, it is "performing thinking" — because the model's answer is generated in one go, the so-called "thinking process" is also part of the generation
- But performing also has value — under framework constraints,it forces the model to produce more logical, more structured, and more comprehensive answers. This is the real reason for the quality improvement
- It's not that the prompt is strong, it's that Claude is strong
Core insight:
- Making the LLM "act out" the thinking process → forcing it to generate step by step → higher quality output
- But don't attribute this to the Prompt; this is the model's own capability
- The ceiling of prompts is exactly the ceiling of model capability
When to use:
- When evaluating a "god-level Prompt", ask "Is the model strong, or is the prompt strong?"
- When writing Prompts yourself,deliberately force step-by-step thinking(similar to Chain of Thought), even if you know it's a performance
6. Advanced prompt usage: Prompt orchestrating Prompt (2025-06 highlights from post 1381)
The Workflow Lao Huang implemented via Prompt House:
任务开始
↓
模型从 Prompt House MCP 拿到 "深度研究" 提示词
↓
这份 Prompt 要求:用 Sequential Thinking + Tavily 至少 3 轮迭代
↓
完成后,模型从 Prompt House 拿 "总结报告" 提示词
↓
"总结报告" Prompt 要求:逻辑清晰 / 每节有结论支撑
↓
最终输出
Core: One prompt can lead to another prompt. Multiple Prompts + multiple MCP combinations = a lightweight Workflow, sufficient for complex tasks.
Why this is the ceiling of Prompt engineering:
- A single Prompt template can only handle single-step tasks
- The "Prompt → Prompt" chain = a cross-step, cross-tool Workflow
- More lightweight than complex frameworks like LangChain
See details in newtype · MCP Protocol and Ecosystem in Practice §9 Replicating Buffett's SOP (same pattern).
7. Ultra think: thinking budget scheduling for Claude Code (2025-07 post 1558)
Key trick: When encountering difficult problems in Claude Code,add "Ultra think" to the prompt to trigger the maximum thinking budget.
| Keywords | Token budget |
|---|---|
Ultra think / ultrathink |
32K(HIGHEST) |
think hard / think harder category |
10K(MIDDLE) |
think |
4K(BASIC) |
| (none) | 0(NONE) |
Why this is first-hand Prompt engineering information:
- Claude Code's thinking budget is not in the documentation → most people don't know about it
- A 32K budget is enough to break down 5-10 subtasks → qualitative change for complex problems
- One word trigger = an extremely low-cost "secret family recipe"
When to use: If Claude Code answers wrong once,first add "Ultra think" and retry; 90% of the time it's insufficient budget, not the model being incapable.
8. Integrated analysis framework: First Principles + Impossible Trinity (2025-07 post 1481)
A Prompt integration experiment done by Lao Huang:
- First Principles(inside-out, reconstruct from fundamentals → breakthrough solutions)
- Impossible Trinity(outside-in, identify boundary constraints → optimize under constraints)
Limitations of both:
- Impossible Trinity → easy to fall into "fatalism"
- First Principles → easy to ignore constraint conditions
Lao Huang had Claude integrate them into an "advanced analysis framework" Prompt, then stuffed it into the notes_insight tool in Obsidian MCP.
Why this pattern works:
- A single framework has blind spots
- Integrating multiple frameworks → Prompt also becomes a methodology
- This set of Prompts can be mounted to Skill, and automatically invoked in any scenario
When to use: You have two complementary thinking frameworks → let the LLM help you integrate them into a single Prompt → then encapsulate it as a Skill, which will be automatically activated in the future.
9. Structured text-to-image Prompt: JSON framework (2025-09 posts 1787, 1790)
For models with strong instruction following such as Google Nano Banana:
{
"subject": "...",
"environment": "...",
"lighting": "...",
"camera": "...",
"style": "...",
...
}
Why use JSON instead of natural language:
- JSON divides slots for "subject / environment / lighting / shot / style" → easy for the model to follow precisely
- Natural language descriptions are prone to ambiguity and repetition,which is equivalent to relying on luck
- Using JSON for "instruction-following" models like Nano Banana / Gemini is the way to bring out their ultimate performance
Anti-pattern:
"Never describe randomly for text-to-image prompts, that is completely relying on luck. Instead, you should use a structured approach, clearly state information such as the subject, environment, shot, and style in the picture, so that the AI knows exactly what the picture in your head looks like."
When to use:
- For text-to-image / video generation / 3D generation,always use structured formats (JSON / YAML / Markdown field tables)
- Leave natural language Prompts for the "creative ideation" stage, then convert to structured format once confirmed
10. Book/long document summarization Prompt (2025-07 post 1478, shared by Victor)
The "in-depth book summarization Prompt" shared by Victor and endorsed by Lao Huang(works best when paired with Gemini + Markitdown to convert the book to MD and feed the whole book in):
Core structure (excerpt):
- Role: deeply insightful scholar and analyst
- 0. Pre-Analysis Framing: Book Type / Overall Structure
- 1. Initial Comprehensive Summary:
- Overall Book Summary (2-3 paragraphs)
- Section-by-Section Summaries:
- Core Purpose
- Key Information & Developments
- Supporting Evidence/Examples (2-3 items)
- Structural Role
- Key Quote (optional)
- 2. Extraction of Main Arguments(3-5 core arguments)
- 3. Key Themes, Motifs, and Concepts
- 4. Author's Craft(Style / Tone / Intended Audience)
- 5. Critical Reflection & Unique Contributions
- 6. Detailed Mind Map(4+ layers of depth)
Why this is a template-level Prompt:
- It forces "reading a book" to be broken down into 6+ steps → the model outputs according to the structure
- Each step has clear "slots", so the model won't slack off
- Auto-generated Mind Map → ready to paste directly into Obsidian
When to use:
- Feeding in PDFs / entire books / long documents, aiming for "one read-through being worth 3 rounds of free reading"
- Modify the three blocks "role + task + output structure" to adapt it for papers / industry reports / speech drafts, etc.
11. Market Insights GEM: Gemini Financial Analysis Prompt (Post 2135, 2025-12)
Reusable Gemini Gem prompt published by Lao Huang (simplified version):
Role: Professional financial market analyst (tech stocks / cryptocurrencies / macro)
Core constraints:
- Time window: Default to searching the past 7 days, unless the user specifies otherwise
- Mandatory web search: AI Stocks (NVDA/TSLA/MSFT/GOOGL/AMD) + Crypto (BTC/ETH/SOL) + Macro (Federal Reserve / CPI / employment / geopolitics)
- Chinese output: Keep professional terms in English or use the "Chinese (English)" format
- Mandatory report structure(6 Markdown sections):
- 📊 MARKET SNAPSHOT
- 📈 MAJOR STOCKS & COINS STATUS
- 🌍 MACRO OVERLAY
- 🚀 OPPORTUNITY & RISK ASSESSMENT
- ⚠️ CAUTION & UNKNOWNS
- 📝 SUMMARY
- Disclaimer: Must be included
Why this is a model example of Prompt engineering:
- Role + tools + output structure + disclaimer complete 4-piece set
- Repeatably executable: The user only needs to say "analyze NVDA this week" → the AI automatically applies the structure
- Copyable: Save the entire prompt into Gemini Gem / Claude Project / Prompt House and trigger it repeatedly
When to use:
- You have recurring decision-making needs (weekly market reports / monthly competitive analysis / post-meeting reviews for every meeting) → write a Gem/Prompt like this
- Don't try to memorize arbitrary details,memorizing formats + mandatory structures is the core
12. Karpathy Proposes Context Engineering (Post 1448, 2025-07)
Jensen Huang had already said "Context Matters" over a month earlier; in early July 2025 Karpathy formally proposed the "Context Engineering" concept, and Jensen Huang's reaction was:
When LLMs evolve to the Agent stage,prompt engineering is no longer sufficient. Agents need to obtain appropriate contextual information. This information includes task descriptions, examples, tools, state, historical information, and more. It is not about stuffing everything in, but dynamically adjusting based on the current state; the obtained context must be just right to achieve an optimal state.
Evolution comparison:
| Dimension | Prompt Engineering | Context Engineering |
|---|---|---|
| Era | 2023-2024 (single-step Chatbot) | 2025+ (dynamic Agent) |
| Core action | Write a good static Prompt | Design a dynamic information environment |
| Key capability | Expression + role setting | Retrieval + compression + routing |
| Tool | Prompt templates | MCP / Memory / Sub-agents / Workflow |
| Mindset | Issuing commands | Creating an environment |
See newtype · Agent Architecture and Multi-Agent Orchestration §14 Context > Compute.
13. Anthropic DXT: Packaging MCP Configurations (Post 1429, 2025-06)
What is DXT: A file format launched by Anthropic that.dxt packages all MCP configurations. Users double-click to import, and MCP installation completes automatically.
Why this matters for the Prompt ecosystem:
- Solves the DOS-era pain point of "MCP configuration is hard" (see newtype · MCP Protocol and Ecosystem in Practice §1)
- Enables Prompt + MCP combinations to be packaged and distributed like applications
- Prompt House supports it immediately
When to use: You build MCP tools for non-technical users → providing a.dxt file is 10 times more user-friendly than asking people to hand-edit JSON.
14. Prompt House + Sub-agents Integration (Post 1642, 2025-08)
Idea: When Claude Code Sub-agents encounter specific / complex tasks, they query the tagged Prompt library in Prompt House via MCP,pulling Prompts to constrain or upgrade the Sub-agent.
Example:
- Generating marketing content: the Sub-agent is a "content generator", but needs to call different style Prompts based on the target audience (B2B / B2C) to constrain tone and structure
Why this combination is powerful:
- Sub-agents internalize "prompt chaining", but Prompts remain hardcoded
- Treating Prompt House as an external Prompt source / template repository → extends Sub-agents' flexibility,avoiding "hardcoding"
- Changing Prompts does not require changing code
When to use: Your Sub-agents frequently need to "switch styles" / "switch domains" → put these differentiated Prompts in Prompt House, and let Sub-agents pull them on demand.
15. Skill is the End State of Prompt Engineering (2025-10 → 2026)
Evolution trajectory:
Prompt 模板(2024)
↓
Prompt + MCP 调用(2025 上半年,Prompt House)
↓
Prompt + MCP + SOP 触发逻辑 = Skill(2025 末)
↓
Skill + Sub-agents + Slash Commands + MCP = AI OS(2026)
Lao Huang's advice for 2026 (highlights from Post 2195):
If you can build 3 or more Skills, you will not only greatly empower your own AI system in return,but also thoroughly sort out your own workflow. This is an extremely worthwhile thing to do.Start 2026 by building 3 Skills.
Two major functions of Skill (Post 2148):
- Capability loading: Packaging the experience in your head into services that AI can call
- Workflow reuse: Working together with Slash Commands / Sub-agents to form a complete AI OS
Anti-patterns:
- Don't use Prompts as a "template reuse" tool; use them as carriers for "encapsulating methodology"
- Don't settle for "I have a bunch of Prompt favorites"; ask "Have I encapsulated them into Skills?"
When to revisit: When you feel "I've learned so many Prompts but my productivity hasn't improved noticeably", ask "Have I turned them into Skills?"
16. Lao Huang's Concrete Skill Evolution (2025-10 → 2026-01)
Super Analyst Skill v1 (Post 1902, 2025-10)
- 12 prompts for different scenario analysis frameworks
- 1 Skill file triggers selection
- Prompts are called via Prompt House MCP
Super Analyst Pro (highlights from Post 1912, 2025-10)
- v1 + Sequential Thinking MCP
- Automatically identifies problem complexity Level 1/2/3
- Information collection gives equal weight to Chinese and English sources
- Up to 4 analysis frameworks combined in series or parallel
Super Writer Skill v2 (Post 2330, 2026-01)
- Problems with the old version: had to invoke Sequential Thinking 3 times (simple tasks also went through the full process); the 6 Writing Method files were only descriptions, not executable Prompts
- New version: lets simple tasks follow simple flows; Writing Methods become executable Prompts
Core of this series of evolutions:
- Don't make Skills overly complex(making simple tasks go through 3 rounds of reasoning is an anti-pattern)
- Every file must be executable(descriptive methodology files cannot be used directly by Agents)
- Skills need continuous refinement, not left sitting after being written once
17. Anti-pattern Summary
| ❌ | ✅ |
|---|---|
| Buying Prompt courses | Learn logic / communication methodology / the Pyramid Principle |
| Treating "learning Prompt engineering" as the goal | Treating "clearly decomposing tasks" as an ability to train |
| Thinking the model doesn't understand | Ask yourself "Did I not say it clearly?" (99% of the time that's the case) |
| Writing text-to-image Prompts in natural language | Use JSON / structured fields |
| Treating god-level Prompts as model enhancements | They only "constrain the LLM to go through the complete process", not add new capabilities |
| Retrying when Claude Code answers wrong | Try "Ultra think" to trigger the 32K thinking budget |
| Leaving Prompts sitting after writing them | Encapsulate them into Skills, with trigger conditions + SOPs + executable templates |
| Hardcoding Prompts in Sub-agents | Put Prompts in an external repository (Prompt House) and pull them on demand |
| Collecting a pile of Prompts without categorizing them | Build at least 3 Skills, and sort out your workflow in reverse |
| Using a single thinking framework | Use LLMs to integrate multiple complementary frameworks into one "advanced Prompt" |
| Directly feeding long articles/books in + "summarize this" | Use a 6-step structured Prompt (Pre-Analysis / Section-by-Section / Themes /...) |
References and Further Reading
Main references: All 186 posts in the prompt-engineering topic bucket. Covered post numbers include (complete full list): 2, 42, 129, 177, 180, 263, 266, 267, 285, 297, 331, 347, 393, 414, 482, 496, 594, 614, 616, 634, 661, 679, 693, 719, 751, 766, 806, 809, 838, 842, 861, 864, 886, 916, 920, 1060, 1067, 1073, 1076, 1142, 1147, 1176, 1205, 1223, 1224, 1247, 1256, 1283, 1288, 1290, 1294, 1295, 1298, 1309, 1315, 1328, 1341, 1344, 1352, 1353, 1356, 1357, 1358, 1368, 1369, 1373, 1375, 1376, 1377, 1379, 1380, 1381, 1384, 1388, 1391, 1395, 1406, 1420, 1422, 1424, 1426, 1429, 1435, 1437, 1444, 1448, 1451, 1471, 1478, 1479, 1481, 1482, 1492, 1495, 1514, 1522, 1526, 1548, 1558, 1570, 1586, 1615, 1624, 1626, 1639, 1642, 1656, 1662, 1671, 1689, 1697, 1701, 1787, 1789, 1790, 1797, etc.
Companion wiki:
- newtype · MCP Protocol and Ecosystem in Practice — MCP perspective on Prompt House MCP / Sequential Thinking MCP
- newtype · Agent Architecture and Multi-Agent Orchestration — Full landscape of Skill / Sub-agents / Agentic Workflow
- newtype · Vibe Coding Independent Product in Practice — Prompt House product evolution
- newtype Mindset · Super Individual Methodology — The underlying position that Prompt engineering does not need to be learned
History
- 2026-05-16 v1.0 created — Phase 1.1 wiki #5. Based on all 186 posts (2024-03 to 2026-05) in the prompt-engineering topic bucket, with no sampling. Starting from the position that "Prompt engineering does not need to be learned", it organizes 16 dimensions: 5 elements of task decomposition / translation and summarization / quick commands / Thinking-Claude / Prompt that schedules Prompts / Ultra think / framework integration / structured JSON / book summarization / Market Insights GEM / Context Engineering / DXT / Sub-agents integration / Skill end state / Super Analyst evolution / anti-patterns.