Step 1: Open Claude Code. Step 2: Create 3 markdown files. Step 3: Type ONE command. You now have an AI team. A researcher that cites every source. A builder that can only use what the researcher found. A critic that never wrote the draft, so it never defends... it. Opus 5.5 runs the whole thing. Hallucinations drop. Quality jumps. You stop babysitting. Full setup with every file inside.show more

Still waiting for the step where someone opens the app and looks at it

Massive stuff

Great point!

Best part is the handoff file. If the builder can only cite what the researcher wrote down, checking the whole run means reading one Markdown file. I make Rho MD, a Markdown notes app.

the critic agent is the real unlock here honestly, stopped me from shipping so much garbage

批评者没写过草稿这点挺关键，少了点护短。

The critic that never wrote the draft is the smart part. It has no reason to defend it.

The critic that never wrote the draft is the key insight — I tried having one agent do everything and it just defended its own bad decisions. Separation of concerns applies to AI teams too. Stealing this setup.

does it actually catch bugs or just nitpick wording?

giving claude code a separate critic agent caught so many of my bugs

i ask gpt to be Claude's critic and vice versa

everyone here loves the critic, but a critic with zero context has its own failure mode: it flags deliberate choices as bugs and the builder "fixes" them. anthropic found ~59% of what their own reviewer kept flagging was old code. so does your critic see only the new diff, or the whole thing?

The weak joint is upstream. If the researcher cites a bad source well, the builder inherits it and the critic checks the draft against that same source. Garbage enters with a citation attached and comes out looking more reliable, not less.

When the critic doesn't trust a source... does it send the researcher back out or does the whole run just stop? A vendor comparison going to finance, say. Every number in it has to trace back to a link

The critic detail is the one I keep coming back to, because my own drafts always get defended by the same mind that wrote them. I tried splitting that job once and the review finally got honest, even if it stung a bit.

Amezing

What if the researcher finds bad sources

its kind of the same thing that happens to people, minus the ego. the writer agent got to its answer through its own steps, so reviewing its own draft just retraces them. a fresh critic starts from somewhere else