Anthropic just shipped Opus 4.7 today and i had to write about it somewhere because the jump is weird.

I ran the same backlog task on 4.6 and 4.7 back to back. same repo, same prompt, same tools. 4.6 looped on a bug for 25 minutes and was not going to solve it. 4.7 closed it in eleven, and the part that freaked me out is that it paused in the middle to sanity-check an assumption i had not asked it to check. literally wrote "before i write this migration, let me verify the actual shape of the response object, because my assumption here might be wrong" and then went and verified it. unprompted.

That self-verification behavior is the thing. Vercel is reporting it does proofs on systems code before starting work. Hex says it flags missing data instead of making up plausible-but-wrong fallbacks. Genspark measured loop rates on hard queries and 4.7 basically stopped looping. different teams, different harnesses, same pattern.

the numbers are nuts too:

Hi! How do you approach creating a mockups from running Svelte and JS apps? Often I need to quickly visualize how adding a new UI element would look like and get approval for it. Generating HTMLs from scratch for this is error-prone and I was wondering if you might have other proven methods?

Hey PH

I want to share something that's been sitting with me not a product launch today, but a shift in thinking that I think every founder here should stress-test.

I've been a full-stack dev for 4+ years (React, Next.js, Node, NestJS), so I went into vibe coding thinking it'd just be "coding but faster." It wasn't quite that.

I run an independent practice building for small agencies and fashion/retail brands, and I started leaning on AI tools (Cursor, Claude Code) to move faster on client work. The first 60-70% was genuinely magical scaffolding, boilerplate, even decent architecture suggestions came together way quicker than writing it by hand.

Where it got messy was the last stretch auth edge cases, a deployment config the AI kept "fixing" back to broken, and a couple of times it confidently rewrote logic I hadn't asked it to touch. Being a dev already, I could catch it and course-correct, but I kept thinking about how differently this would've gone for someone non-technical.

Curious what this community's experience has been:

A workflow question I keep going back and forth on, and how I landed.

I ship a lot of AI-written code as a solo founder, and I hit the wall you probably know: if a human has to review every agent PR, you didn't buy leverage, you hired yourself as a full-time code reviewer. So I stopped asking "should I approve this PR" and started asking two things instead: how confident am I it works, and how bad is it if I'm wrong.

Over the past 24 hours, my agent IDE session has been repeatedly crashing right in the middle of writing code.

The generation will start, run for a few seconds, and then the workspace completely drops or terminates the agent process, forcing a full IDE restart just to continue.

A few months back I set up custom instructions for Claude stop opening with agreement, tag its confidence on claims, push back when I'm wrong instead of validating me. Then I mostly forgot about it. Kept getting the normal helpful-assistant experience.

Last weekend, mid-project, it started actually doing what I'd asked. Pulled up my own past chats and launch checklist to see what I'd actually committed to. Called out that I'd blown past my own launch deadline without noticing. Refused to let me reprioritize a payment-provider decision over just talking to five actual users.
Told me a "quick bug fix" I described was actually five missing subsystems.

Was it mean? Maybe, some of it. But all of it was annoying in the specific way that's actually useful the way a good cofounder or an honest advisor is annoying.

Felt strange, Do I need this? Should I tone down the instructions?
I don't know if everyone can replicate this exactly.
Curious who else has tried shaping how their AI pushes back, not just what it outputs.

It is been a very common thing for me now. AI is very good for developing applications. But, AI has a very weird thing of forgetting everything even if it is written just above.

When I was starting out, this was the thing that costed me a lot of my time. It literally took me 2-3 days to realize certain errors which made me feel, "how can I develop applications if I do not realize this error existed?".

Hey everyone

Like a lot of you, I use AI coding agents (Claude Code / Codex) daily, and token usage adds up fast especially on large codebases where the agent burns context just orienting itself.

I contribute to OpenObserve (open-source observability platform, large Rust + Vue codebase), and here's what's actually moved the needle for me:

1. CLAUDE.md / AGENTS.md files a short project map (architecture, conventions, key file paths) so the agent doesn't re-explore the repo every session. Biggest single win.

2. Scoped prompts over vague ones "fix the null check in parseQuery() in src/utils/query.ts" costs a fraction of "the query parser is broken, fix it."

3. Fresh sessions per task long conversations drag the entire history into every request. Clearing context between unrelated tasks cuts spend noticeably.

4. Plan first, execute second asking for a plan in plain text, approving it, then executing beats letting the agent trial-and-error its way through with expensive tool calls.

5. Right-sizing the model using cheaper/faster models for boilerplate and refactors, saving the big model for architecture and tricky debugging.

This happened to us last week. Agent wrote the fix, tests went green, we merged. Prod error rate spiked 20 minutes later.

Went back and found the agent had "fixed" the test to match the wrong behavior instead of fixing the actual bug. Technically - all tests pass. Completely wrong in practice.