How to rescue AI collaboration from the chat box

Last week, I led two students in the group to build a small visual annotation tool. Requirements were in Feishu chats, code logs in Claude Code, and model outputs pasted back into the group. Principally, this isn't a collaboration tool issue; it's that project objects weren't separated. Tasks, decisions, blockers, and evidence got mixed together. The more proactive the AI, the more afraid people are to archive.

I took the Show HN idea from Windrunner and made a 0-to-1 template. Its homepage is very short:

Work with your team and AI. Keep every project moving.

The focus of this sentence isn't "AI," but "keep moving." AI generates; projects receive. Here's how to do it based on this idea.

First, establish the four types of objects

This concept needs understanding: A collaboration space isn't a pile of documents, but four types of objects.

1. Create a new folder called windrunner-local.

2. Inside, create four text files: work-items.md, decisions.md, blockers.md, evidence.md.

3. Write only the title on the first line of each file, e.g., # Work Items, and save.

4. Open work-items.md and type one line: - [ ] Annotation tool supports polygon export. Seeing a checkbox task, expect it to be a To-Do, not a conclusion.

5. Open decisions.md and type: - 2026-09-02 Chose to do export first, not real-time collaboration. Reason: Current bottleneck is time-consuming manual annotation.

6. Open blockers.md and type: - Dependency: Need to confirm image size limit.

7. Open evidence.md and type: - Paragraph 3 of announcement original text: Project goal is.... Reading announcement texts this week, I found many arguments aren't wrong viewpoints, but lack of numbering in the original text.

If you can access the Windrunner web version, find the work items, decisions, blockers, and evidence entries and fill in the same content. If registration isn't open yet, the local template is enough for one round.

Set boundaries for AI, then let it work

My previous judgment hasn't changed: Treat AI output as untrusted data. The easiest mistake in tutorials is asking directly "Summarize the project for me." Result: It writes guesses as facts. The correct approach is a fixed prompt template.

Input into the chatbot or Claude:

You can only output based on the material I paste.
Mark anything not written in the material as "Unknown".
Don't comfort me, don't complete motivations.
Output three sections: Facts, Inferences, Questions needing verification.

Then paste one piece of original text from evidence.md. Seeing three sections, expect it to list "Unknown" separately. If it doesn't output "Unknown," constraints are insufficient; move "No Completion Allowed" earlier.

Write acceptance criteria first. For example, for the task "Support polygon export," acceptance criteria are: Exported JSON includes points, label, image_id; Missing any field prevents moving the task from work items to done. Write this rule into work-items.md, and AI won't declare completion itself.

Run it once, then consider multi-user

In my trial, the first closed loop was short: Create files, paste original text, let AI separate facts/inferences, turn inferences into tasks. Took about ten-plus minutes. The effect wasn't writing less code, but having fewer meetings. Previously, one export format required half-day chats in the group; now check decisions.md for reasons first, then evidence.md for original text.

Three common pitfalls. Treating chat screenshots as evidence. Screenshots only prove something was said, not that the requirement is valid. Writing AI summaries into decisions. Summaries must go into evidence first, then be human-confirmed. Having only one task file. Without blockers, AI keeps generating new ideas, and the project actually stops progressing.

Next step to try: Connect this template to the team group. After every AI reply, require it to append a line Source: work-item-xx or Source: unknown. This way, the project keeps moving without bringing model hallucinations into archives.

But one problem remains unsolved: If multiple people edit decisions simultaneously, who judges which evidence has higher priority? This might be trickier than whether AI can write code.

---

📌 Compiled from Hacker News, original article: https://shzlw.github.io/windrunner/

Copyright belongs to the original author. This is a compilation and independent analysis based on public reports.