Stef

15+ yrs in SWE / SRE, now building AI SaaS & autonomous software. Creator of SSVE (governed AI meta-harness). Solo builder turning vision into real products.

[Jurly @jurlycat]

Claude writes ~90% of Claude Code’s code. Agents now do most of the code review, too.

The Claude Code team at Anthropic just published how their review process works:

• Before a PR opens: runnable checks, then a second agent reviews the diff with fresh context.
• On every PR: parallel agents find bugs, and separate agents vote on whether each finding is real.
• Security review and CI gate merging.
• Low-risk changes from trusted authors can merge without human approval. Higher-risk changes need one or two people.
• Merge queues, feature flags and staged releases help contain mistakes.

From June to September, they report merged PR volume more than doubled while the share needing a revert fell by over 40%.

But there’s a great detail buried in the post: one PR went through 50+ rounds of agent review and fixes overnight.

They now ask the author to reconsider the approach after three rounds, focus later reviews on the new diff, and suppress repeated findings.

Writing code got cheaper. Keeping two agents from inventing an endless workload for each other is now an engineering problem.

[Stef @shipwithstef]

@finebutwhy @poteto @bot Over bot or with external harness?

[Stef @shipwithstef]

@poteto @bot All cool but I need 4.8 to feel more deterministic. :D

[Stef @shipwithstef]

@SimonasLTU1 Yep, I think they are slowly taking away some of the weekly but still looked good.

[Stef @shipwithstef]

@SPAC89 To be honest this looks to be old one or I am beta testing for a while now, not too bad for extra usage.

[Stef @shipwithstef]

@IndependentEco @grok what is the current state of 200$, is it still same usage as 100$ but just higher 5h limits or it changes since 5.5 era?

[Stef @shipwithstef]

@bcherny Do you recommend specific setup.

[bolt.new @boltdotnew]

Bolt is free for EVERYONE. For the ENTIRE weekend.

Starting at 12:00am PT October 17.
Build it Saturday, ship it Sunday, submit by Oct 20 for a shot at $10K.

Ready to build? Details here 👉 https://t.co/vzI8Nngl7G

[Stef @shipwithstef]

People are you happy with new grok ultra subscription, any feedback will be appreciated - meanwhile if this turn out to be dead post please research for such talks as this is pretty recent @grok

[Stef @shipwithstef]

@ZryMiller Wait what, do I read it out correct - message from another session resend the entire context not just the message?

[monokern @monokern]

SpaceXAI's Grok Bot team put their own bot templates in the marketplace, and this is a f*c*ing goldmine

you can import a setup built for a specific job, connect your tools and adapt it to your own work

here are 10 worth looking at:

bot 1 → dr eggbot designs and creates other Grok Bots. describe the job you need covered and it helps build a specialist for it.

bot 2 → Lingxi's Engineer Bot supervises cloud coding agents and watches their PRs. you handle the merge.

bot 3 → Stalk Bot watches competitors and reports changes to their products, pricing and positioning, with sources attached.

bot 4 → Haggle Bot finds wasted SaaS spend and drafts vendor negotiations for your review.

bot 5 → Proto Bot turns feature ideas into prototypes inside your actual app, delivered as draft PRs.

bot 6 → Video Editor picks takes from raw footage and builds captioned vertical videos, with an edit plan you approve first.

bot 7 → figma bro turns Figma frames into build specs and checks your components and design tokens.

bot 8 → Distro Bot works on your marketing campaigns and creative, with approval before spending, posting or sending.

bot 9 → Startup QA Bot walks through your product in a test account and reports bugs with screenshots and steps to reproduce them.

bot 10 → Signal Prospector finds buying signals in your CRM and product data, then drafts outreach around the evidence.

worth looking through before spending another evening writing a giant system prompt from scratch

[Stef @shipwithstef]

@kamilbaranek_ As long as this bring me real value and especially if that help me make more money I can subscribe to many more.

[Stef @shipwithstef]

@tirthaexe dunno, I believe there was announcement about this for Opus -> @grok afaik since 5.5 for Opus specifically they said there is no cache reset on effort switch, please fact check

[Stef @shipwithstef]

Effort switching reset cache again for Opus spotted. I just moved high to medium and got prompted - if true that is huge nerf. Hope it was a glitch anyone faced similar fate?

[Stef @shipwithstef]

@neerajjj6785 claude ops 5.5 high right now

[Stef @shipwithstef]

@eVshevtsoV pretty much living in the harness you mentioned, what's up

[Stef @shipwithstef]

The Instagram post is only the 250$ offer and does it include anything else from the other offers or all are decoupled? Regardless do you have some data/feedback on people that have used it I think I am hooked to explore launch something important through it soon, will do some checks on my side first but it looks like a crazy deal in terms of algorithm.

[universal-modder @universalmodhq]

3d gen "models" are pipelines: shape diffusion → mesh → remesh → uv unwrap → texture → bake. each vendor packs different steps.

[Stef @shipwithstef]

No but, I did grok this one out and the result was fascinating, your whole existence now resonates with agentic harnesses all over the world, joke a side no idea what this is but what it say it do is pretty on the spot -> In the AI coding agents / vibecoding world, “gm” has a second, specific meaning.

It refers to a skill + MCP server (and “canonical universal harness”) for coding agents like Claude Code, Cursor, Gemini CLI, and Codex. A gate decides when work is done—not the agent. It keeps a plan of tasks, runs real checks, and refuses to let the agent mark anything finished until the evidence and predicates pass.⁠GitHub

This matches the ethos of people building with AGI-style coding agents (including CobaltHere and Cobalt Cockpit): agents don’t get to unilaterally declare “done.” Verification, state machines, CI markers, clean worktrees, and recorded evidence take priority over pure vibe acceptance.

So the original “Do we still do GM?” post works as a double entendre—crypto Twitter “Good Morning” plus the agent-harness “gm” that enforces real completion gates instead of free-form vibecoding

[Stef @shipwithstef]

@kamilbaranek_ no