Three Common Failure Scenarios First

Writing changelogs and release announcements for SaaS products sounds like the least technical work there is. But browse through various products' changelogs and you'll find plenty of train wrecks.

First type: the technical perspective hijacking the user perspective. Developers casually paste in commit messages—"refactored the order state machine," "optimized the cache eviction strategy." The writer thinks it's perfectly clear; users finish reading completely confused: so what does this have to do with me? Is my problem solved?

Second type: batching three months into one session. When a small team gets busy, the changelog is the first thing sacrificed. By the time a pile of features accumulates and you try to backfill, the details are long forgotten, and you end up staring at diff files guessing "what was this change for again?" The resulting announcements are dry and stiff—users can't sense that the product is alive.

Third type: having AI copy directly. In the past couple of years, many teams have started using AI to write these, but incorrectly: they dump the entire git log in and say "write this into a changelog." The result is AI output that's just translated commit messages—"fixed several bugs, improved performance and stability." Users have seen that kind of language too many times—it says nothing at all.

The root of the problem isn't the tool, it's the process. Here's the process I run now—one that an indie developer or a team of three to five people can adopt directly.

The Right Process: Feed AI "User-Perspective" Raw Material

The core reason AI writes bad changelogs is that the input contains no user-perspective information. My process has four steps:

Step one: build a "semi-finished repository" of change records. I use a shared doc and require the team (including myself) to jot down three lines each time a feature is merged: what changed, why it changed, and which user scenario benefits. The notes are allowed to be rough—AI will clean them up later. This habit costs two minutes each time.

Step two: process everything together before release. On release day, I give the AI the semi-finished repository content, some key git logs, and the high-frequency user feedback from the past two weeks in the community group. The user feedback part is critical—it determines the narrative order of the announcement: whatever users care about most goes first.

Step three: constrain the output with a structured prompt. Don't just say "help me write it"—explicitly tell the AI who the reader is, what the tone should be, how to structure it, and that every item must answer "what does this mean for the user." My template is at the end of this post, ready to copy.

Step four: do a ten-minute human final review. Focus on two things: whether the AI has hallucinated feature details (this is the biggest pitfall—AI occasionally writes "planned" features as "already shipped"), and whether the release date and version number are accurate. Ten minutes is enough.

Before and After Using AI

Efficiency: Previously, writing a formal release announcement took two to three hours from digging through records to final draft—and because it was painful, it always got procrastinated. Now the raw material accumulates daily, and AI processing plus human review keeps the whole thing under half an hour. More importantly, the release cadence went from "whenever we remember" to a fixed biweekly schedule.

Quality: Announcements used to be a list of "what we did"; now they're a narrative of "your problems now have solutions." For the same database optimization, we used to write "optimized query performance"; now we write "project lists with thousands of records now load noticeably faster—users who previously reported slow loading, give it another try." Completely different information density.

User communication: The announcement explicitly ties back to users' earlier feedback, essentially telling users "we heard your suggestions." This feeling of being seen has a bigger impact on retention than the features themselves. Of course, I won't make up specific numbers—you can A/B test the two writing styles yourself.

A Ready-to-Copy Prompt Template

你是一个SaaS产品的更新日志撰写助手。请根据我提供的材料，生成本次更新公告。

# 背景
- 产品名称：{产品名}
- 产品定位：{一句话描述，如"面向小团队的工时管理工具"}
- 版本号：{版本号}
- 发布日期：{日期}

# 材料
- 变更记录（含改动原因和场景）：
{粘贴半成品记录}
- 相关的用户反馈摘要：
{粘贴用户群里的问题/建议，没有可写"无"}

# 写作要求
1. 读者是产品的实际使用者，不是工程师。禁止出现类名、函数名、内部模块名。
2. 每条更新必须包含两部分：做了什么（一句话），对用户意味着什么（具体到使用场景）。
3. 如果材料中的用户反馈与某条更新相关，在该条中明确呼应，如"之前反馈过XX的同学可以关注这条"。
4. 顺序按用户价值排序，而不是按开发工作量排序。
5. 语气：专业但有人味，可以用一两句轻描淡写的幽默，不用感叹号堆砌，不夸大（"革命性""颠覆"这类词禁止出现）。
6. 只使用材料中出现的事实，不得补充任何未提及的功能或数据。
7. 输出结构：一段30字以内的整体摘要 + 分条更新列表（每条不超过3行）+ 一句结尾（引导反馈渠道）。

The last requirement is a lifesaver—AI is most prone to fabricating feature details when "freestyle writing," and explicitly forbidding this cuts the final review time from half an hour to ten minutes.

A Few Additional Tips

The cost is negligible. This kind of text-processing task works perfectly fine with mainstream cost-effective models—no need for flagship models. For specific choices and pricing, refer to each provider's official pricing page. If you're already using an API aggregation service, these high-frequency small tasks are especially well-suited to a unified gateway, saving you from opening a separate account just to write announcements.

Write the announcement and changelog separately. The changelog is a complete list; the announcement picks highlights and tells a story. You can first have AI generate the full changelog, then ask it to "pick the three to five items most noticeable to users" and expand them into the announcement. One set of raw material, two outputs.

Accumulate your tone samples. After two or three months, pick a few human-edited final drafts and add them to the prompt as style references. AI output will increasingly sound like "written by your own team," and final-review edits will keep shrinking.

Writing changelogs is, at its core, translating the developer perspective into the user perspective. AI happens to be best at exactly this kind of "translation"—provided you give it the right raw material and clear constraints. Replace the placeholders in the template with your product info, and you can try a version today.

If you don't yet have a go-to model API, check out this aggregation gateway—register and start using it right away, with multi-model switching that's perfect for the combo of "cheap models for daily small tasks, flagship models for complex final-review scenarios": https://api.thistoken.ai/register

---

Tired of juggling provider integrations? Register at https://api.thistoken.ai/register and call every model through one base_url.