June Retrospective: Turning AI Workflows Into a Delivery System

TL;DR: June changed how I think about AI workflows. Writing, publishing, notes, product updates, and feedback started to connect into a system that can deliver, review, and keep improving. If you are using AI for content, products, or a one-person company, the useful lesson is simple: keep real inputs, make tools readable by agents, turn knowledge bases into products, use new platforms as demand signals, and let publishing feedback shape the next decision.

---

At the beginning of June, my expectation for AI workflows was still simple.

I wanted AI to help me write faster, publish more reliably, and spend less attention on formatting, copying, organizing, and archiving.

By the end of the month, the real change was no longer about speed.

I worked on several connected things: organizing flomo notes, upgrading md2wechat into a more agent-friendly tool, expanding the Feishu version of the Jieni AI Practice Manual, testing WorkBuddy as a new acquisition channel, and sharpening my thinking about industry Agents from “better search” to “capturing expert judgment paths.”

Those things may look separate, but they point to the same question:

How can one person turn AI from a temporary assistant into a system that keeps delivering?

My answer is getting clearer. Do not only ask whether the model can write. Ask where the inputs come from, how judgment is preserved, whether tools can be called correctly by agents, whether content becomes a product, and whether feedback returns to the next decision.

§1. Keep Real Inputs First

In June, I processed another batch of flomo notes.

In the first half of the month, I synced 19 files and 42 notes. Near the end of the month, I synced another 14 files and 25 notes. I ignored the 2022 and 2025 historical notes and only handled the 2026 material.

This reminded me that many systems do not start with a beautiful knowledge base. They start when you are willing to preserve real inputs.

Real inputs are often messy.

They may be a voice note, a user question, a judgment from a conversation, a product friction point, a platform signal, or a thought you have not fully understood yet.

If those things are not saved, your agent can only read public material and the context you give it in the moment. It becomes an outsourced writer, not a system that understands your long-term judgment.

After June, I became more convinced that the first value of a note system is not classification. Its first value is preserving the judgment scene so that your future self and future agents can read it again.

This applies to anyone building an AI workflow.

You do not need a complex directory structure on day one. Start with one place where you keep three kinds of material:

1. ·Problems you keep running into.
2. ·Questions users, customers, or readers keep asking.
3. ·Judgments that feel valuable even when they are not fully formed.

Three months later, those notes will be more useful than asking AI to invent ten topics from nothing.

§2. Tools Need to Be Readable by Agents

md2wechat changed quickly in June.

I focused on the capabilities around v2.7.0, v2.8.0, and v2.9.0:

1. ·Fixed egress and WeChat IP whitelist support, which solves the last-mile problem of creating WeChat drafts from changing network environments.
2. ·Agent image plan mode, where tools like Codex, ChatGPT, and WorkBuddy can generate images themselves after md2wechat provides the prompt and format.
3. ·WeChat title suggestion, where md2wechat reads the article, prepares a title-generation request, and lets the host agent or external model continue.

On the surface, these are product updates. Underneath, they reflect a change in how I think about agent tools.

When building a CLI for humans, the main question is whether a person can use it.

When building an agent-native tool, you also need to ask:

1. ·Can the agent discover the right command?
2. ·Does the command return enough information for the next step?
3. ·Which actions are safe preparation, and which actions have side effects?

For example, title suggest should not quietly rewrite the title or create a draft. It should prepare the title request and make the next step clear to both the user and the agent.

Image plan mode should not pretend an image already exists. It should tell the agent what the image is for, what aspect ratio it needs, and what prompt to use. After the image file is generated, then upload and draft creation can happen.

That boundary matters.

The longer an AI workflow becomes, the more dangerous it is to blur every step into “auto complete.” Stable systems make inputs, outputs, and risks inspectable.

If you are building an AI tool, check it with this question: does it only help humans click buttons, or can an agent understand, call, correct, and continue the workflow?

§3. A Knowledge Base Should Become a Product

I spent a lot of June updating the Jieni AI Practice Manual.

The md2wechat section grew to 22 pages. I later added three more pages:

1. ·Using md2wechat inside WorkBuddy to write WeChat articles.
2. ·The five-layer structure of a content growth system.
3. ·Industry Agent case studies, from retail research to intelligent cooking.

The point was not just moving articles into Feishu.

I increasingly do not want a knowledge base to become an article warehouse. An article warehouse makes the author feel organized, but the reader still does not know what to do next.

A Feishu manual should feel like a delivery product.

It should answer:

1. ·Where should a beginner start?
2. ·Which page helps when configuration breaks?
3. ·What should be checked first when something fails?
4. ·Which prompt or instruction can be copied for a real scenario?
5. ·What should the reader do after finishing the page?

That is why I split md2wechat into pages for versions, installation, configuration, WeChat drafts, images, title suggestions, and WorkBuddy scenarios.

Reducing product friction is not only about features. It is also about whether a reader can follow the documentation and finish the first path.

The same is true for creators.

If you have written many articles, try turning them into a small manual. Do not organize it by publish date. Organize it by reader task.

Readers do not come to inspect your output history. They come with a problem.

§4. Content Growth Needs Feedback Loops

My view of content growth also changed in June.

Previously, I cared most about moving an idea to publication: draft, title, format, publish.

Now I break the content growth system into five layers:

1. ·Input: flomo, Get Note, voice notes, user questions, platform feedback.
2. ·Resources: old articles, Feishu manuals, community reviews, FAQs, hot topics, historical questions.
3. ·Writing: main question, brief, draft, voice check, fact check.
4. ·Distribution: WeChat, X, paid community, Feishu, jieni.ai.
5. ·Feedback: saves, reposts, comments, DMs, consulting questions, purchases, and the next topic.

The fifth layer is the one I most want to improve.

Without feedback, a content system optimizes by feeling.

With feedback, the system can see which articles created reader action, which titles only sounded good to the author, which directions deserve more writing, and which directions should be downgraded to short posts.

This is what I kept realizing while working on X growth, WeChat growth, and Feishu manual updates in June.

Publishing is not the finish line.

Publishing sends a judgment into the real world so that it can receive feedback.

That is why I care more about what happens after publishing: archive the final version, record the URL, save platform feedback, review the individual post, and move stable judgment back into the wiki or manual.

If you want AI to help with content, do not only ask it to write the next article.

Ask it:

1. ·Which pieces created real interaction in the last month?
2. ·Which questions kept repeating?
3. ·Which pieces deserve to become tutorials?
4. ·Which directions should stop getting attention?
5. ·Which validated judgment should the next article inherit?

That is closer to growth than asking for ten more titles.

§5. New Platforms and Old Industries Are Demand Sources

Two lines of thinking influenced me a lot in June.

The first was WorkBuddy.

I noticed that some users discovered md2wechat through WorkBuddy tutorials. That signal matters. When a new agent platform starts rising, many users do not yet know what real work it can do. A tutorial that completes a real task builds trust faster than an abstract introduction.

WeChat writing is a good WorkBuddy scenario.

The user may not search for “WeChat formatting tool.” They may search for “how to write WeChat articles with WorkBuddy,” “how to create a WeChat writing expert,” or “how to send an article to the draft box.”

That means the new platform itself becomes a distribution surface.

If you are building a tool, do not only wait for users to search for the exact category name. Watch what users are trying to learn on new platforms, then embed your tool inside that task.

The second line was industry Agents.

After talking with a friend who works in retail supply chain consulting, I understood industry Agents more concretely.

General search returns many pages. An industry expert knows what to inspect first: store distribution by city, hiring roles, company registration, map heat, official accounts, operational traces, and user reviews.

Each signal is weak alone. Together, they help judge whether a brand is expanding, shrinking, testing, or only creating noise.

Intelligent cooking has a similar lesson.

Generating recipes is not the hard part. The hard part is closing the loop across temperature, vision recognition, ingredient ratios, washing and cooling, cookware state, and output consistency.

That made one thing clear to me: the point of an industry Agent is to productize expert judgment paths. Web search is only one step.

For programmers and AI creators, this is an opportunity.

Do not stay only inside the AI tool circle and compete on model update speed. Look at real workflows in traditional industries: hiring, customer service, sales delivery, store operations, content acquisition, and report generation.

Many jobs that can be amplified by agents will not show up as “AI Agent demand.” They hide in budgets, roles, complaints, consultations, and repeated information work.

§Five Reminders From June

If I compress the month into five reminders, they are these:

First, input matters more than inspiration.

Recording real problems consistently is more reliable than asking AI to invent topics on demand.

Second, tools need to be readable by agents.

The clearer the command, return value, boundary, and side effect, the more stable the workflow.

Third, organize knowledge bases by reader task.

Readers do not need your output history. They need to know what to do next.

Fourth, feedback must return after publishing.

Without feedback, a content system only produces. With feedback, it starts learning.

Fifth, demand often appears in new platforms and old industries.

New platforms create new entry points. Old industries hold real budgets. AI creators should watch both.

§How to Borrow This in July

If you want to build your own AI workflow in July, keep the first version small.

Start with this:

1. ·Choose one input place and keep real thoughts and user questions there.
2. ·Choose one delivery scenario you repeat often, such as writing WeChat articles, making reports, posting on X, or organizing customer questions.
3. ·Break that scenario into fixed steps. Decide which steps can be handled by agents and which steps need human confirmation.
4. ·Turn the result into a tutorial that a reader can follow.
5. ·After publishing, record the feedback and decide what to write next.

One month later, you will have two things.

More stable output.

Clearer demand judgment.

That is more valuable than chasing another tool.

The biggest lesson I took from June is this: AI is not only useful when it helps you finish one task. It becomes much more useful when every task leaves a trace and improves the next decision.