Articles, templates, and postmortems without the growth detours

73 Building Your Own Personal Agent Recently I've noticed that many people are quietly building their own personal assistants. There are several schools of thought that are quite interesting, so I'll share them with everyone.

1. The Tool-Person School The human is the AI's camera, recorder, and robotic arm. The human helps the AI solve the context problem in exchange for the AI's intelligent experience. For example, I buy a full set of equipment, wear smart glasses all the time, turn on video and audio recording, and regularly upload it to my AI so it knows about my daily life, my work content, everything that happens to me every day... After accumulating data for a few days, the AI has enough context to understand me quite well. At that point, I just need to add a few prompt words guiding it to proactively help me, and the AI can remind me that I'm about to be late, that too many clothes have piled up and it's time to do laundry, that my girlfriend seems to have something on her mind today, and suggest I leave work early to spend time with her... Trading context for an intelligent experience is like this—giving up some...

Is 72 REA actually useful? An AI tool that went viral in October, claiming it can reverse engineer and analyze anything. That made me a bit curious, so I put it to the test and will report the conclusions directly: the results, the principles, the test conclusions, and my judgment.

1. Results It took 1.5 hours to reverse engineer and analyze with GPT 6.1 Sol max, then using DeepSeek flash to fully replicate the original Mac Fans Control App was done in one 1.5-hour round, and the core functions work. REA reverse engineering analysis produces 4 documents: a reverse engineering report, functional specifications, an independent implementation architecture and development roadmap, and acceptance test cases. With complete documentation, implementation can be handed directly to AI, which is equivalent to AI giving requirements to AI, very comfortable. 3 hours doesn't look complicated, right? I used qwen 3.8 m...

71 "AI Programming 2.0" RIPER5 2.0: The RE Protocol Last week, the RIPER5 protocol I had used for 2 years was officially retired on my end Countless contributions and hard work, laden with honors Let me think about a question: it's already 2026 and GPT 5.6, do we still need any Rules? Is it perfectly fine to just run codex naked? Not fine. The reason is that the model is very strong, and codex's Harness is also very strong, but people are weak The fact is that people who use AI are not good at expressing themselves; the boss makes a casual decision with one sentence, and the people below run their legs off—this familiar plot repeats itself on AI every day So the real problem now is that AI is more rule-abiding than people; people are too unruly and too lazy; people need to abide by rules and also need to be lazy This is the original intention behind my upgrade of the RIPER5 protocol I thought about removing it, but I was afraid the AI programming results would get worse, so among these three articles, the most...

70 "AI Programming 2.0" RIPER5 Is Getting Old

Something that's good to use must be something that makes people lazier

If you can sit, you don't stand; if you can lie down, you don't squat—that's human nature, and only efficiency tools that go along with human nature will make people willing to use them

The previous post summarized why the RIPER5 protocol works so well: because of the 6 key roles it plays

Although RIPER5 is very good, the models and Coding Agents have also developed over these past few years, and in the latest AI Coding tools, RIPER5 has indeed caused some trouble

AI Coding has never been a one-trick-pony solution that works everywhere, and I also have to admit that RIPER5 is getting old

Where is it old? Let's take a look together

At present, frontier LLM coding capabilities are very strong and have fully reached Agentic Coding. When using RIPER5 in daily work with frontier models, I've encountered...

69 Why "AI Programming 2.0" RIPER5 Works

I recently finished something big: upgrading the RIPER5 protocol to 2.0. Friends who previously bought the programming secret manual have all given this protocol rave reviews, and I myself have used it for more than 2 years. Along the way, I kept pondering a question: why can this ancient protocol from the Claude 3.7 era instantly beat a whole crowd of dazzling SDD approaches and work surprisingly well? I have finally summarized it clearly recently, and I am sharing it with everyone.

Note: The original RIPER5 first came from a Cursor community expert. I did not change a single word and used it for 2 years; it really works. On the surface, this protocol has these effects: RESEARCH: delay action, obtain repository facts, expose information gaps INNOVATE: force divergence, avoid prematurely locking onto the first solution PLAN: decompose tasks, externalize thinking, establish shared expectations between human and machine Chec...