Made with RiseMore.

Lilach Bullock has a rule I can't stop thinking about.

Cancel any tool you haven't opened in two weeks.

She's a solo productivity writer who spent years testing every AI tool that launched. Her conclusion: most tools don't remove a step — they put a prettier interface on the same manual grind. She calls it the Removal Test.

Her stack: ChatGPT Team, Claude Pro, Fathom, Zapier, Descript. Five tools. Under £150/month. Replaced what would have been an £800/month human VA.

The tools she ditched are the interesting part. Motion calendar — "just a prettier interface on the same manual scheduling problem." Superhuman email — "made me feel productive without actually removing decisions." Notion AI — "added AI to a tool I already had instead of removing a step."

It's the same test I keep running against RiseMore.

AI writes your copy fine. It won't tell you what's worth writing about. Most tools schedule the post — you still have to draft it. They track competitors — you still have to turn that into action. Figure out what matters, draft it in your voice, publish where it counts: that's three decisions strung together, and it's one too many when you're also shipping product, answering support emails, and fixing the bug someone just found.

RiseMore removes the chain. Not just one link.

The "cancel anything not opened in two weeks" rule hits different when you're the one paying for 14 subscriptions out of your own pocket. Software bloat is a tax solo founders pay in silence. Every tool you don't use is a decision you haven't unmade.

What would pass the Removal Test in your stack today?

Investors now check AI-code provenance: can your team explain what the AI built, or does only the LLM understand it? The term is Vibe Debt. Same problem in validation: if you can't trace where a claim came from, you're vibe-validating.

Saw a Product Hunt review yesterday that named the real problem with AI marketing agents — and then solved the wrong one.

Michael Vavilov on a new "Approval-first Agent OS": *"CMO decisions I'd trust least — audience choice, messaging and distribution have slow noisy signals that make agent learning hard."*

He's right about the noise. Marketing signals are slow. An AI running on generic prompts can't tell if your messaging flopped or the right people just haven't seen it yet.

But the objection misses the actual architecture question.

The gap isn't signal clarity. It's whether the agent starts from zero — blank prompt box, no memory — or from your actual product context, positioning, materials, and what you've already tried.

Those are completely different tools.

The Soloop team describes their CMO agent as something that "surfaces the patterns and brings you the signal so you can make the call in minutes instead of weeks." Directed execution. Human judgment at the approval gate. That's not a magic box — that's an operator who actually knows your product.

I think the reviewer wanted: describe your company once, agent handles the rest forever. That's not how it works. Probably shouldn't be.

The trust layer isn't smarter AI. It's what the agent knows before it starts writing.

That's the whole problem I'm building RiseMore around. Not generation — grounding. Because "AI writes your copy" is solved. "AI knows what's worth writing about, for your product, for your audience" — that's still the hard part.

Waitlists feel like traction. They're not.

Ken Rutkowski calls it a Demand Mirage — converting early noise into payroll before real demand exists.

Nobody pulled out a credit card. That's the whole test.

Speed without distribution isn't momentum. It's just faster obscurity.

That's the thing no one says clearly about the vibe coding wave: AI made building faster, but it didn't touch the second mountain. You can ship in a weekend now. You can still watch it land in silence for months.

"Vibe marketing" is the term starting to crystallize around what comes next — an AI-native operating model where founders provide strategy, taste, and judgment, and agents handle execution. Not AI as a writing assistant. AI as the layer that keeps distribution running while you're heads-down on the product.

The framework needs two things to work: customer data and a brand context layer. Without both, AI just produces the wrong thing faster. Generic content, generic voice, generic results. The speed multiplier becomes a liability.

Microsoft flagged this shift in their own leadership data — 844 of 9,037 senior leaders already operating in agent-led ways. The role isn't tool operator anymore. It's agent manager.

For a solo founder, that reframe matters. You're not trying to hire a $150K marketing hire or grind out a content calendar at midnight. You're trying to stay in the build loop without letting distribution collapse every time you go heads-down.

That's the gap RiseMore is built for. The context layer that makes the execution sound like you — not a template, not everyone else's AI output — while you direct instead of do.

Still early. Still building in public. But the architecture is real.

What's breaking your distribution loop right now?

#vibemarketing #solofounder #indiehacker #buildinpublic

Don't cheerlead.

A founder built an AI decision partner on that rule. Its job: surface the failure modes you've missed.

Most founders do the opposite — ask "do you like my idea?" and count every yes as proof.

If feedback can't say "this won't work," it isn't validation.

AI agent 不出错的时候，你根本看不到它在漂移

没有报错，没有异常，只是悄悄走错了方向

所以设计 agent 的第一原则不是「它能做什么」 而是「它失败时会怎样」

法院已经判了：Air Canada 要为 chatbot 的错误信息负责

你的 agent 准备好被问责了吗？

微软开源了两个 AI 安全工具：RAMPART + Clarity

RAMPART：把对抗性攻击写成 CI/CD 测试，像测 bug 一样测安全 Clarity：在写代码之前，先找出设计层面的安全漏洞

核心思路： 安全不是上线前的检查 是贯穿整个开发流程的工程纪律

对建 AI agent 的人来说，值得看一眼

HN 上一个讨论让我有点不舒服

AI agent 用你的 OAuth token 开 PR、评论、approve 代码 审计日志显示：是人做的

但实际上没有人看过这段代码

AI review 了 AI 写的代码

这不是效率问题，是问责问题

Google I/O 最低调的一个发布：WebMCP

浏览器原生 AI agent 标准 Chrome 149 已有 origin trial 让 agent 直接执行 JS 函数、填 HTML 表单

不需要 MCP server，不需要 Playwright

如果这个标准跑通，web agent 的门槛会低一个数量级

同一个功能，用两个模型分别检查，比单模型 code review 强很多

Claude 4.7 Max 写代码 Codex GPT 5.5 做 review——更「轴」，专揪边界情况 Claude 再修 最后新开 session，两个模型从头看 staged changes

防上下文偏见

2026 年 vibe coding 的反面长这样

If an AI validation tool always returns a verdict, I assume the verdict is manufactured — and I'll trust the ones I like.

Real trust comes from the tools willing to say 'not enough evidence yet.