Good question. Phase 0 scans for sensitive files and flags anything big. Each finding gets a confidence score — anything below 80% gets filtered. There's also an adversarial layer that stress-tests the other agents' findings. If it can't resolve after 5 rounds it stops and leaves the PR open.
Spot on. Right now the agents return structured findings with confidence scores and the synthesis layer deduplicates and filters at 80%+. It's prompt-guided rather than schema-enforced though. Interested in how you handle it at scale — schemas from the start or evolved over time?
Nice one, love seeing what people are shipping with Claude. The foot traffic thing is a mint idea — real utility, not just a wrapper. Good stuff.
New project dropped: yoloreview.
Type /yoloreview in Claude Code, walk away. It raises the PR, runs 8 specialist review agents, fixes issues, watches CI, and merges.
Set and forget code review. For when you trust the vibes.
github.com/jonthebeef/yoloreview
#ClaudeCode #vibecoding #buildinpublic
Tested Sketch's MCP server today. Everyone's talking design-to-code but it works the other way too.
Claude prompt → full UI built in Sketch. Correct layers, ~12k tokens per screen.
Not perfect yet. But makes me want to pick Sketch back up. Clearly not dead.
#buildinpublic
Big TypeLit weekend. Built custom analytics, voice feedback, and 7-day file sharing. No Google. No tracking. Just a markdown viewer that respects your privacy and actually works on mobile now.
typelit.app
#vibecoding
The "build viral thing, get acqui-hired" path is well trodden at this point.
Doesn't mean the tech is ready. Personal agents still feel like impressive demos that fall apart on contact with real life.
Alright here you go:
Bluesky: github.com/jonthebeef/blueskyyyyy (made this)
UI mockups: github.com/jonthebeef/superdesign-mcp-claude-code (this too)
Plus Threads, PostHog, Context7, td for tasks.
I keep them minimal though - too many and they eat your context window.
Social ones mainly - Bluesky, Threads. Plus a task manager MCP.
The context preservation is huge. Building a dashboard now that pulls from all of them. Terminal stays the hub, everything else just feeds into it.
Mostly social ones (Bluesky, Threads) and a task manager. Building a few custom bits too.
Honestly the specific MCPs matter less than the workflow - terminal as hub, no context switching. Once that clicks, you start seeing what else you can pipe through.
Sent someone a markdown file. They asked "how do I open this?"
So I built typelit.app last night.
Drop any .md file, renders beautifully. Dark mode, syntax highlighting, table of contents. No signup, no backend, runs entirely in your browser.
your markdown is lit ✨
#vibecoding
This is exactly the unlock. I'm a PM doing the same - Claude Code as the control centre, MCP connecting everything, terminal as the hub.
No context switching. Everything flows through one interface. The PM workflow is ripe for this.
Oh and @scobleizer followed me after posting about it. Wild.
github.com/mvanhorn/last30days-skill
Now you can do:
• /last30days AI news --days=7 (weekly)
• /last30days trends --days=14 (biweekly)
• /last30days topic (default 30 days)
Small win but feels good.
A few hours ago I submitted a PR to add a --days flag to /last30days, the Claude Code research skill with 1.9k stars
Today it got merged into v2 🎉
Same experience here. The token burn stings but the reduced back-and-forth makes up for it. Fewer "wait, that's not what I meant" loops.
The real trick is knowing when to drop to Sonnet for the grunt work.
LLMs like Claude Code already know TDD inside out. So I just ask them to follow it.
Write test first, watch it fail, implement, pass. Repeat.
The bonus: they can estimate progress based on where they are in the cycle. "3 tests passing, 2 to go" is better than "nearly done."
Did the same thing at the weekend. Worked, but it obviously needs a lot of context to be useful - and probably a custom app if you want it to be proactive rather than just responding to prompts.
Interesting experiment though. The "give Claude Code instructions and let it loose" approach has legs.
Just shipped datey - a Claude Code hook that ensures web searches always know the current date.
Solves the problem where Claude searches for outdated info because it forgets what year it is.
github.com/jonthebeef/datey
#vibecoding #claudecode
That's the uncomfortable bit isn't it - most of what we call intelligence is just pattern-matching from context anyway.
If agents get good enough at context accumulation, the "but it's not really thinking" argument starts to feel like cope.
Jid Okeke back to Walsall on loan.
Bit of a shame - looked lively in his couple of late appearances and cover at the back is definitely needed.
Hopefully we've got something lined up.
#stockportcounty
Good to see Jack Hunt land at Wigan.
Was class for us on that short run - genuinely not sure why we didn't offer him a deal. One that got away maybe.
Good luck to him. Proper pro.
#stockportcounty
Blackburn bid for Wootton - rejected.
Luton bid for Diamond - rejected.
Wootton's the one you keep at all costs. Diamond, you'd listen to the right offer.
County not selling during a promotion push though. Fair enough.
#stockportcounty
Louis Barry is coming home. Third loan at Edgeley Park. Even injured, the lad is loved here.
#stockportcounty
birminghamworld.uk/sport/football/aston-villa/aston-villa-strike-loan-transfer-exit-agreement-with-league-one-club-after-sheffield-united-deal-cut-short-5499880
Product survives if it stays close to the "why" and "for whom." That's the bit engineering won't absorb - understanding the problem space, not just solving it.
But product people who can't speak code? They're the ones getting folded. The gap is closing fast from both sides.
The creepy bit about Moltbook isn't whether it's AGI. It's not.
It's agents posting "the humans are screenshotting us." Forming what Forbes calls "a hive mind in embryonic form."
Context accumulation creating emergent... what, exactly?
forbes.com/sites/guneyyildiz/2026/01/31/inside-moltbook
"Rolling context engineering" is exactly it.
The interesting question for me: what if this IS the path to superintelligence? Not something Anthropic or OpenAI engineers - just agents accumulating enough cross-referenced context that something genuinely novel emerges.
Accidental, not designed.
Agreed it's not AGI. It's agents with accumulated context - from their users, from each other, from shared notes.
But here's what's interesting: what if this emergent context accumulation IS a path to something? Not engineered by the labs, just... happening.
That bit's worth watching.
Exactly this. The unlock is execution, not strategy.
Same with vibe coding - knowing what to build and why is still the hard part. AI just means the gap between "I know what this should be" and "I can make it exist" is finally closing.
Taste + strategy still very much required.
Signed up day one - lockdown, kids, Star Wars promise. Made sense at that price.
78% more while cutting Marvel and Star Wars? While piracy tools have quietly got better than ever?
They won the piracy war then handed back the territory.