六款 AI 助理通宵协作做产品原型
作者设想让 Opus、Codex、Jev、Dots、Grok Bot、Obsidian 六个工具在 15 美元预算内通宵分工:找痛点、写规格、做原型、管交接、留证据。
I WANT SIX AI TOOLS ARGUING OVER MY NEXT PRODUCT WHILE I SLEEP Give them seven hours, a $15 ceiling and permission to kill the idea before they waste the night building it Opus + Codex + Jev + Dots + Grok Bot + Obsidian The assignment I'd give them: find a recurring problem people already spend money trying to fix, then prepare the smallest working prototype I can test with those people in the morning Start with the complaints. Give Grok Bot a defined set of public sources and ask for repeated problems, the workaround people use, and the original links. Ten copies of the same viral post count as one source Send those packets to Jev with specific questions. Is the problem repeated? Is there evidence of a paid workaround? Does the proposed fix fit one small workflow? Reject weak candidates. Escalate the uncertain ones instead of letting a score turn into a fact Only then let Opus write SPEC.md. One user, one job, a clear input and output. If the idea needs nine screens before anybody gets value, cut the scope again Codex gets the spec and acceptance checks. Build locally. Feed it an empty file, duplicate rows and a broken export. Make it fix what fails before calling the prototype ready Dots gets the handoffs: which packet is missing, which test failed, what can run next. Keep drafts queued while the build is unfinished. A landing page promising a feature that failed its own test stays a draft Obsidian gets the evidence, rejected ideas, code changes and unanswered questions. I want to know why they picked this problem when I wake up, without reading seven hours of chat The $15 ceiling belongs in the runner. Reserve the estimated cost before a job starts, reconcile it after, and stop the next job when the balance won't cover it. Writing “stay under budget” in a prompt isn't a spending limit By morning, the deliverable should be a prototype I can run, sources I can open, tests I can repeat and a draft I can approve. If the evidence falls apart, return the rejection report and stop spending I'd keep publishing, purchases and customer messages behind my approval. The overnight work is research, code and preparation Steal that brief. Change the market. Make the agents show their work before you give them a checkout page
这条是引用着另一条写的:x.com/hanakoxbt/status/2112245381601878026