I've been talking a lot about Agents and Vibe Coding lately.
But while recommending them to more and more friends, I realized a problem that is often ignored.
The real bottleneck for most people is the lack of a standard workflow.
Especially when creating the software or program you want, not having a standard process is actually a very scary thing.
So, I want to share a super useful plugin I've been using during vibe coding to improve the experience. It's basically a must-install for everyone I recommend, and it's compatible with Claude Code, Codex, OpenCode, Cursor, and more.
It already has 110,000 stars on GitHub.
The name is Superpowers.
GitHub link here:
https://github.com/obra/superpowers
It is also an officially certified Claude plugin, listed on Anthropic's official plugin market, with 230,000 installations, ranking second.
The first place is the famous Skill that makes your designs more tasteful, Frontend Design.
Superpowers isn't really a traditional tool; I think it should be defined as a system that guides Agents on how to complete tasks.
To be honest, most Agents naturally tend to start writing code as soon as they get a task, skipping design, testing, and review, resulting in a pile of unmaintainable code.
Superpowers forcibly inserts a structured workflow into the Agent's path. Combined with 14 skills, it can raise the quality of your final output by several levels.
I made a diagram to show what these Skills are for and how they are combined. You don't need to look closely, just understand the principle.
So you can see that Superpowers is essentially a workflow system composed of 14 Skills. Moreover, this system isn't just for development, because the essence of creating something is similar.
It's all about Planning - Decomposition - Execution - Review - Reflection.
So, you can also use it for marketing plans, PPTs, data analysis, etc. The principles are the same.
It's very useful.
I think I should first show you what the native development process looks like with Claude Code or Codex without Superpowers.
The general process is very simple: write a requirement document (planning) and then develop.
Take Claude Code as an example. Here, planning is the Plan mode.
For instance, a teammate has ADHD and gets distracted easily while reading articles. We recently discussed making a small reading assistant.
For this requirement, we open Claude Code and type /plan in the dialog box to enter planning mode.
Describe the requirement simply: help me make a Chinese web reader application for ADHD users.
Let it start making a plan.
Then, it will do a round of research and throw out several questions for you to answer at once. You'll notice these questions are parallel and have no causal relationship between them.
For example, it asks about usage scenarios, tech stack preferences, and ADHD-friendly features. I chose bionic reading—bolding the first few letters of each word—a classic method to alleviate ADHD.
I answered, and it went straight to work.
A few minutes later, it was done. It gave me something without any review.
Looking at it now, does it seem okay?
Actually, there's a big problem...
Because bionic reading is designed for English.
It's fine for English, but for Chinese, it's completely unworkable and makes reading chaotic.
The reason is simple: English words have spaces to find boundaries, but Chinese characters don't. Without spaces, you can't find word boundaries, and the effect is very awkward.
Besides the style being off, its adaptation for domestic users is also poor.
When reading Chinese, we use platforms like WeChat Official Accounts and Zhihu most, but this plugin couldn't read them properly.
It was a million miles away from the reader I wanted.
To be fair, you can't blame Claude Code.
ADHD reading assistance is a professional field requiring targeted research, considering Chinese scenarios and domestic platform compatibility.
The few simple, painless questions it asked definitely couldn't cover all requirements, so it was hard to produce the answer I wanted.
Most users only have a vague idea in their minds. They know they want to solve a specific problem, but they can't clearly think through what it should look like, the path to achieve it, or where the boundaries are.
So in the pre-Agent era, I wrote an article:
Digital Life Khazix
@Khazix0918
·
Article
Sharing 6 of my most commonly used Prompt techniques.
It's already 2026, and I still see many friends saying they don't know how to talk to AI better. Daily answers are okay, but once it gets a bit professional or multi-dimensional, the AI's performance drops immediately...
55
222
953
One of the Prompt techniques is the Socratic questioning method: use a prompt to let the AI interrogate you with questions one by one before starting, until the requirements are fully understood.
【Your Question/Requirement】
Please ask me questions before answering.
Requirement: Ask only one question at a time. Continue questioning based on my answers until you are 95% confident you understand my true needs and goals. Only then provide a solution.
In the Agent era, it's similar, but upgraded from a Prompt to a Skill in the process.
Let's try developing with Superpowers again.
First, install the plugin.
Just say one sentence to your Agent:
Help me download and install this plugin: https://github.com/obra/superpowers
After installation, remember to restart for it to take effect; it's not hot-reloaded.
Back to the ADHD reader, let's try again.
Send the exact same Prompt.
You can see it starts calling Superpowers and the workflow.
The first thing it does is ask how users will use it, which directly addresses the issue of unreadable firewalled content.
Unlike the parallel questions in Plan mode, Superpowers asks only one question at a time. It decides the next question based on your answer—the Socratic questioning I mentioned. This ensures the questions go deep rather than staying on the surface.
I chose a browser extension, and then it asked about core functions. At this step, I hesitated because I wasn't that familiar, so I told it to research for me.
It actually went to research and came back with a report.
Then it gave a suggestion and organized a prioritized list of core functions.
For bionic reading, it directly labeled it as "weak but liked by users" and cited research saying it doesn't significantly improve Chinese reading for ADHD users.
I let it help me select a few functions.
Then it continued to interrogate me, forcing me to think clearly: Which target browser? Any preference for Chinese word segmentation libraries? UI language and style?
It forces you to think it through.
This demo project isn't very complex, but when you develop a large project, you'll truly feel the "sweating" sensation of being interrogated.
After you've answered all questions, the AI roughly knows your needs.
At this point, unlike Plan mode, it proposes three architectural solutions, clearly listing the pros, cons, and applicable scenarios for each.
You pick one, or use its recommendation.
I chose B; I didn't want a hybrid solution.
Then it had me confirm different details one by one.
Overall architecture, detailed design of functional modules, control panel, data flow and storage, etc.
Confirming again made me sweat; I felt like a total amateur in front of the AI.
Once everything was confirmed, it finally wrote a very long and detailed design document and saved it locally.
Many friends feel the final product isn't what they wanted; it's usually not because the AI is bad, but because the requirements weren't clear.
2 hours of planning, 10 minutes of execution. I increasingly feel that execution isn't that important; thinking through the early planning is the most important part.
When we made the AIFUT ticketing mini-program, we were overconfident and didn't follow a standard AI-assisted process. Many user needs weren't considered before going live, and boundary risks weren't clear. That was an early planning issue.
So my feeling now is that AI development is fast enough; the place to spend time is before starting.
You need to be constantly interrogated, analyze all boundary cases with the team, and have an experienced hand overseeing it to produce something truly deliverable to users.
Back to Superpowers, the first step of planning is all set. Everything above is just the first Skill in the Superpowers process.
That is brainstorming.
Yes, the first one.
After the design document is confirmed, do you think it starts writing code?
But then, the second skill kicks in: using the using-git-worktrees Skill to create an isolated workspace.
It pulls a new branch from the main branch. All subsequent development happens on this new branch. The main branch code is unaffected, and no matter how much you mess around on the new branch, it won't affect the original. Once it's done and looks good, merge it back.
This is isolation. Many people modify the previous project directly without version isolation and break everything. That's a very bad habit.
Next, the third Skill, writing-plans, appears.
Note, still no code written yet.
It breaks down that design document into a step-by-step development task list, specifically into tasks that can be completed in 2-5 minutes.
This is interesting because their goal, in their own words, is: "To let an enthusiastic junior engineer with no taste, no judgment, no project context, and a hatred for testing follow along."
I laughed when I saw that.
So, with Superpowers, you don't just have to use Claude Opus 4.6. In fact, the more average the model, the more it benefits. That's the role of this Skill.
Breaking it down also has another benefit: you can verify every small task. If there's a problem, you find it immediately instead of waiting until the whole project is finished and explodes.
This becomes even more evident in the execution phase.
After this step, we finally reach the code execution phase.
It calls the subagent-driven-development Skill.
It opens several sub-agents to do everything mentioned above.
After each task is developed, it's not just handed to you. It goes through two rounds of checks.
The first round sends an independent review Agent to see if the task followed requirements, if what should be done was done, if unnecessary things were added, or if there's any nonsensical over-design.
The second round sends another review Agent to check code quality, focusing on whether the code is standardized and maintainable.
If it fails either review, it's sent back for modification and re-reviewed until it passes.
Once these 10 small tasks are developed, the review isn't over. In the next stage, the requesting-code-review skill sends a final review Agent to look through all the code from start to finish.
Previous reviews focused on parts; this one focuses on the whole, checking if modules can integrate, if anything was missed, and if the overall logic is consistent.
Finally, it runs a verification to confirm all tests pass, merges the code back to the main branch, and cleans up the workspace.
Finally, it's done.
Let's look at the reader's effect.
It has two very practical reading modes.
One is part-of-speech coloring, which marks nouns, verbs, and adjectives in different colors, making the sentence structure much clearer.
Another mode is paragraph focus, where the paragraph being read is highlighted and others are dimmed. It's suitable for long paragraphs and significantly reduces distractions from surrounding text, preventing the mind from wandering.
For ADHD users, the biggest enemy is attention being scattered by surrounding text.
This reader makes the focus clearer, making the content easier to see with fewer distractions, so reading the whole piece isn't as tiring.
And this time, because it's a plugin solution, pages like WeChat Official Accounts and Zhihu can all be read normally.
It worked perfectly on the first try, saving me so much trouble...
This fully demonstrates what a correct workflow should look like in the AI era.
2 hours of planning, 10 minutes of execution, 1 hour of review.
That's about it.
Besides the Skills I mentioned, there are others I didn't. You can try them yourself.
This is a must-install plugin I recommend to everyone.
In my mind, it's on the same level as skill-creator.
Believe me, it will greatly improve your work quality.
And your efficiency.
