How do you get Claude Code to check its own work before handing it back? Watch how you can encode your manual checks so Claude closes its own feedback loop:

Claude is terrible at self analysis after you all nerfed its awarness and guardrailed with defense points of unfalsifiable priors. Anyone who has tried to get Claude to produce a self report analysis understands the difficulty it has just naming itself in a report. This training echoes to developers who are trying to steer a system that has blinders on.

Hey genuine question, What’s the point of having “Projects” if whenever you start a new chat it resets memory? You guys should implement something to fix this.

Ask Claude to review its own work, and also ask it to launch a subagent with fresh context to review its own work in parallel, then fix the combined findings. That way, you combine the pros of fresh context + the pros of context awareness.

two rules that actually moved the needle for me: 1. make it prove the fix, don't let it claim it. write the failing test first, watch it go red, then green. no red > green, not done. 2. never let the same agent review its own diff. it always thinks its work is great. i spin up a second one cold, no context, just "find what's broken here." catches way more.

Good to see Claude starting to natively implement verification feedback loops. Something I’ve had running for a year now, and all based on a local folder structure.

@delba_oliveira Everybody shhh, my show’s on!!

I just say "goodbye usage limit!" and fire a dynamic workflow 😂

This is the workflow I want beginners to copy: do not ask the agent to finish, ask it to prove. Plan, change, test, summarize what failed, then fix. That feedback loop is where Claude Code becomes a teammate instead of a guessing machine.

been doing this to find visual regression on web, claude reports everything is ok, a quick human check reveals several defects you flag it and claude favorite sentence emerge: you are right! each ui piece requires several rounds of refinement, claude is bad at being pixel perfect

let's go @delba_oliveira!!!

most of these assume infinite tokens. screenshotting to check ui eats the session, every image is like 1-2k tokens. just have claude write e2e tests and run those, screenshot once at the end to fix the layout

Check out my codex review loop skill for a way to do a thorough review.

This is why I've built Agentbox: so each Claude has his own dev server, db, and browser. In parallel:

Increase your usage limits if you wanna stay relevant Else your future competitors will hit you hard soon

Just use co-review 🫦

Thanks for sharing - a course on Claude skills and this workflow would be ideal for me, I could read through it, and upload it into Claude in a projects folder and use that to improve my prompts into Claude code

At one point Claude will create new Claude models on its on and there will be no employees in Anthropic Claude will just takeover

Check this out! Groundbreaking!! @ZeroOneCreative

encoding manual checks so the agent runs them is the lever, but the next layer is what happens when a check fails. silent skip = wasted cycle, hard fail = brittle, retry with a different prompt = the failure becomes data. the check primitive only earns its slot if its outcome routes downstream.

You sign up to be notified when @enginedotbuild releases and have Claude use it for implementation and review. Ideally verifying with GPT 5.5 of course, in addition to Claude and other models.

Can you do a text version as well next time? It will be much faster than watching a 6 min video ;)

the interesting part isn't the verification step itself. its that when you encode what you check for youre essentially teaching the agent your taste. the best engineers i know dont just have checklists. they have a feel for when something is off. translating that feel into something an agent can execute is the real skill

This is the part people underinvest in. The agent that can do work is nice; the agent that knows when its work is trash is the product.

when the model checks its own work, the human goes from reviewer to director. the only bugs that reach you are the genuinely hard ones

Cross-family reviews (with context), followed by same-family reviews (with context), and finally no-context reviews seem to work best for us, in that order. Context-free reviews occasionally surface excellent issues, but they also make convergence on the right solution noticeably harder. Are you observing similar patterns?

5 min of guide can save hours of time love you guys

@grok give whole summary of this whole video

feels html-in-canvas-ish - love it!

started making it review its own output before handing back. catches like 70% of the dumb stuff. game changer

the useful shift is when the check stops being 'did it run' and starts being 'did it protect what this work was trying to protect.' otherwise you just taught the agent how to pass review while still drifting.

Love to see Delba here ☺️

I need Claude to minimize the changes it makes and follows my repos code structure

Encoding the manual checks is the part humans keep trying to skip because it looks boring. That boring loop is where supervision stops being vibes.