返回首页

对比阅读

对比阅读:AI Going Off Track? Move Your Work Context Too 与 多人共用 AI 还总跑偏?先把工作上下文一起搬过去

AEN
AntigravityAI 编程自由职业·

AI Going Off Track? Move Your Work Context Too

The scariest part for me wasn't that AI couldn't answer

At 10 p.m. last Tuesday in Hangzhou, I was helping a three-person squad improve its quoting workflow: one person organized client requirements, another drafted replies, and another checked pricing. But every time we moved the AI records around, the answer seemed to be missing half a page of background. I got stuck too: after two hours, I realized we'd all been working from three different sets of documents.

If you've also had to re-explain the project to every person and every tool, maybe it's worth looking at Antigravity's new /teamwork command. Rather than training another smarter model, it lets AI divide the work across a shared team space, with tasks, code, and progress visible to everyone.

What it actually solves is context handoffs

Antigravity is Google's intelligent development environment. In plain terms, it puts coding, running projects, and AI-assisted changes in the same workbench. The new feature lets team members launch parallel tasks through /teamwork: one person builds a page, another adds tests, and a third perspective reviews the result.

Chen Yu, a developer-tool maker in Beijing, recently tried it while rebuilding a personal product. At 10:30 a.m., he assigned old-page migration, API verification, and test coverage to three collaboration roles. By noon, the team had a gap list to work from, rather than three conflicting sets of edits. The name is fictional, but this relay-style workflow, where people compare differences while working, closely reflects how small teams actually operate.

Recreating it today: low budget, but not a one-click solution

On paper, this research feature currently has no separate price, and public materials haven't revealed its long-term cost. For now, I treat it as 0 yuan in cash. The real time cost is 2–4 hours: one hour for one member to organize the requirements, then another 1–3 hours to split the tasks and verify the results. The technical barrier isn't high, but I still need to understand basic project files, software versions, and testing workflows.

The first step isn't to explore every feature. It's to open Antigravity, click “New Team” or the team entry point, and give two roles a small page change to collaborate on. Button names may vary by version. I wouldn't load the whole project into the first test; I'd start with a small task deliverable in half a day.

Keep the setup restrained, based on your stage

Just starting: If I'm running a solo side project, I probably don't need multi-person collaboration yet. Keeping requirements and version notes in one place offers better value. If I do try it, I'd limit the test to two AI roles reviewing one small feature.

Serving 1–2 clients: If every delivery means repeating the background, I'd start with a pilot combining client requirement breakdown and reply drafting, then check whether shared context genuinely reduces rework.

Scaling up: If I'm leading a 3–5-person squad, I can put task allocation, result checks, and handoff notes in the same space. I'd still keep human approval in place and avoid having multiple roles edit the same thing simultaneously.

Not everyone needs this tool, and even mature teams may not want it connected straight to core projects. It's fine not to try it now. When context handoffs really start slowing me down, testing it with one small task is the cheapest way to find out.

BZH
AntigravityAI 编程自由职业·

多人共用 AI 还总跑偏?先把工作上下文一起搬过去

今天最让我犯怵的,不是 AI 不会答

上周二晚上十点,我在杭州给一个三人小队改报价流程:一个人整理客户需求,一个人写回复初稿,另一个人核价。可 AI 里的记录换来换去,答案总像少了半页背景。我也卡过:忙了两个小时才发现,大家用的是三份不一样的资料。

如果你也遇到过“每个人都得重新解释一遍项目”的情况,或许能看看 Antigravity 刚加入的 /teamwork。它的思路不是再训练一个更聪明的模型,而是让 AI 在一个团队空间里分工协作,共享任务、代码和进度。

它解决的,其实是“上下文交接”

Antigravity 是 Google 的智能开发环境,简单说,就是把写代码、运行项目和让 AI 协助修改放进同一个工作台。新功能允许团队成员通过 /teamwork 发起并行任务,例如一个人做页面,一个人补测试,再由第三个视角检查结果。

北京做开发者工具的陈屿,最近在一个个人产品重构里试过:上午十点半,他把旧页面迁移、接口核对和测试补充分给三个协作角色;中午十二点,团队先看到的是一份差距清单,而不是三份互相冲突的修改稿。名字是虚构的,但这种多人接力、边做边看差异的流程很接近实际小团队用法。

今天复刻:预算不高,但别把它当一键按钮

账面上,这项研究功能目前没有单独标价,公开材料也没给出长期价格;可以先按 0 元现金成本理解。真正的时间成本是 2—4 小时:先用 1 小时让一位成员整理需求,再花 1—3 小时拆分任务、核对结果。技术门槛不算高,但至少要懂基本的项目文件、软件版本和测试流程。

第一步不是研究所有功能,而是打开 Antigravity,在团队空间里点 “New Team”或团队入口,把一个很小的页面改动交给两个角色协作;具体按钮名称可能随版本变化。第一次别塞整个项目,先拿一个半天能交付的小任务测试。

按你的阶段,用法可以很克制

刚起步:如果你是个人副业玩家,现阶段可能不需要多人协作,先把需求和版本说明集中放好更划算;想试,也可以只让两个 AI 角色互审一个小功能。

有 1—2 个客户:如果每次交付都像在重讲背景,我会先拿“客户需求拆解+回复初稿”做试点,确认共享内容是否真的减少返工。

在扩规模:如果你正带 3—5 人小队,可以把任务分配、结果检查和交接记录放进同一空间,但保留人工审批,别让多个角色同时改同一处。

这工具不是所有人都需要,成熟团队也未必适合直接接入核心项目。现在不试也没事;等你真的被“上下文交接”拖慢,再拿一个小任务验证,成本最低。