... sessions in one repo, one killed a dev server it thought was its own. The fix that actually stuck wasnt another line in CLAUDE.md, it was giving each session its own worktree and port range so they physically cant touch each others stuff. Prompts are advice, isolation is a rule Comment from Justin Oberbauer(@usa374): @neelagiri_chettiyar   @gfazioli  Thanks I have also been trying to figure out this workflow. I normally work on a local mac, but want ... ... actual task a dependency bumped here, a config default flipped there, nothing that fails a test. I only caught one later because a build behaved differently on a clean machine. Now I make agents list every file they touched

... improve" with, it can only pick from what already exists. Second, I name the boundary in the request, "change only the button row in this file, nothing else", and I review the diff, not the screenshot. A diff that touches ten files for a one line ask is my signal to roll back and ask again. Is your design system in code already, or mostly in Figma for now? Comment from Michael Hurley(@michaelh_rley): One thing that helps ... ... solo maker: keep a short 'do not touch

... because it thought it looked cleaner. The original bug got fixed, but the extra refactoring broke a completely unrelated feature that i only noticed days later through a user email. Now i have to explicitly tell it not to touch other files in every single prompt. Do you guys use any specific system Rules or prompt constraints to keep agents from drifting? Comment from Shivam Kushwaha(@shivam_kushwaha16): AI coding agents can be unpredictable with scope once they start making ...

... thousands of generated labels subtly non-compliant. So it is frozen. Bugs elsewhere get fixed; that folder gets read-only respect. Do you have a part of your product like this? Something you have consciously decided is done, where touching it is banned no matter how much it itches? Comment from Siarhei Hamanovich(@hamanovich): In my case it is the column names in my database. We renamed a pair of fields across the whole codebase because the old names read ...

... data. Different people need different levels of access. Your teammates need to understand how the app works so they can change it too. And the app needs to keep evolving without turning into a black box nobody wants to touch. At that point, would you be comfortable letting AI change anything it wants? We ve been thinking about this a lot while building WeWeb MCP (launching on PH tomorrow!) Agents like Claude, ChatGPT, Cursor, and Gemini can now build apps ... ... making sure the builder stays in control of both the app and the AI building it. That s why we built guardrails into the way AI works inside WeWeb: Before AI builds: choose what AI can and can t touch

... Synapse is an AI agent execution guardrail. It runs as a persistent local engine in memory with sub-millisecond query latency. The core difference isn't just speed: Synapse simulates edits in memory and validates them before they ever touch disk. Graphify can tell you who calls a function . Synapse can tell you whether it's safe to delete that function with compiler-level confidence scores, blast-radius risk levels, and a definitive safety verdict. One gives your agent ... ... list. The other gives your agent a decision. 2. The Linux Kernel Benchmark (74,619 files) We cloned torvalds/linux locally and ran cold index extraction and queries on both engines: Filesystem Discovery: Synapse 2.0s vs

... Could use some help/feedback in exchange for free letters to your loved ones : ) Here are the details and the problem I keep running into .. I mean to keep in touch with my grandparents but life gets in the way and it's always on the back burner. Texting is pretty useless bc they aren't great at using their iphones, calls have a 10% hit rate, and "I'll write a letter this weekend" never comes to fruition ... ... with the photos built into the page, and mail it to your grandparent's door. $14/month for one recipient (less than a Hulu subscription) and instead of you watching something....you get the good feeling of actually staying in touch

... your answer is honestly 100%, you re a better driver. Genuinely. If it s not, and you touch the phone even once in a while, that s the habit I think Undivided can help with. Only if you want it to, though. See how many hands-free miles you ve earned, keep a record of them, and unlock the collection you deserve for leaving your phone alone and improve driving habits ...

... noticed a pattern with AI coding agents that bothered me more than the actual code quality. I'd ask for a small change. The change would work. Then I'd look at the diff and realize the agent had touched: 11 files for what I thought was a 2-file task. Sometimes there was a good reason. Sometimes it was just architecture drift disguised as cleanup. So I started experimenting with a simple rule: Every task gets a change budget ... ... Before implementation, I ask the agent to identify: files it expects to modify files it expects to add dependencies it expects to touch

... carry the weight people will want to put on it. Would you want your own writing marked because you asked a model to fix your commas? Comment from Hossein Yazdi(@hosseinyazdi): Exactly. There s a huge difference between AI touched this text and AI authored this text. The former feels much easier to detect, but also much less useful ...

... idea behind it. The plan creates a cheap checkpoint. I've also started treating the assumptions section as especially important. If the agent says: "I'm assuming X already exists" that's something I can verify before letting it touch the code. It turns the workflow from: Prompt Code Review into: Prompt Plan Challenge assumptions Code Verify For me, that small change has made AI-assisted development feel much less like "letting the agent run" and more like actually collaborating ...

... Since agents will happily rebuild anything you ask in minutes, "I'm afraid to touch this" stops meaning what it used to. But I don't think the fear goes away, it just moves. I keep a section in my CLAUDE.md that isn't architecture notes, it's just "already tried this, here's why it didn't work." Without it the agent will quietly rebuild the same mistake in 3 months, because it has no memory of the first ...

... wrong today, what is the largest number of people who find out before you do? For a coding agent that is usually one. You, reading a diff. The moment the agent sends, books, charges, replies to a customer or touches production, that number stops being one, and most of us have never actually worked out what it is. So whether you run agents or ship them to other people, which of these are you? A. Nothing my agent does reaches ... ... third case it still misses and it is the one that got me. Some of the worst reach comes from actions that are trivially reversible AND technically correct. Writing a number into a notes file is fully undoable. It touches

... three hundred milliseconds for repeat visits. This is the kind of speed that simply cannot be achieved with traditional, heavy database-driven websites without spending thousands of dollars on enterprise-grade enterprise-level application delivery networks. --- Day 6: Mobile Touch Target and Accessibility Adjustments On Saturday, I spent my morning testing the mobile experience across several physical devices, including an older Android phone and a couple of iPads. Mobile optimization is about much more than just responsive grid layout ... ... about physical usability. People in medical environments are often moving quickly, wearing gloves, or multitasking, so touch

... AVE could already analyze footage locally and distinguish what was happening in different clips but the asset library could still ...

... here about AI unlocking new products and speeding everything up, but I'm curious where the community sees it going the other way... where "AI-powered" is actually a turn-off for users, or where a human touch is the actual selling point. One example I keep thinking about: curated recommendations. Best movies, books, restaurants, coffee shops. There's so much AI-generated recommendation content out there now that a human-picked list (with a name and a point ...

... Been going back and forth on this. Right now my setup lets Claude Code run migrations and seed scripts directly against a local dev database, but I've kept it firewalled off from ever touching staging or prod, even read-only. Everything that reaches prod goes through a PR I review manually. Saw a comment in another thread here about a subagent going off-script and targeting its own orchestrator, which made me second-guess even the local access ...

... weeks in, and I still haven't touched a desktop. Everything architecture decisions, debugging, database migrations, the whole thing happens through a phone screen and an AI agent doing the actual typing. What's strange is how different the failure modes are from what I expected. I thought the hard part would be not knowing how to code. It wasn't. The hard part was learning to think in prompts instead of logic how to describe a bug precisely enough ...

... reset the database and took the data with it. Nothing malicious, that is what the command is for. After that, `db push` and `migrate` are banned outright in that project's config, a `pg_dump` is mandatory before anything touches the database, and we moved most projects onto Drizzle with hand-written numbered SQL migrations that a deploy script applies. Nothing in the chain now generates destructive DDL on its own, which means the agent is no longer the thing ... ... happens more than you think! Comment from Vishal Rastogi(@maganuk): For me it was an agent that had write access to more than it needed, and it helpfully cleaned up a config it was never asked to touch

... function in one of my projects that's been "temporary" for about 4 months now. It works, it's covered by tests, and it's ugly enough that every time an agent touches that file it asks if I want it cleaned up. Every time I say yes, I regret it a little. Not because the refactor is bad, usually it's genuinely cleaner, but because now I've burned review time on code that wasn't broken, and there ... ... small chance the agent introduces a subtle behavior change I won't catch until it's in prod. So now my rule is basically: only refactor if I'm already touching

... code. Here is what I actually use it for. 1. Full account audits in minutes. Feed it raw data. Get a client-ready brief back. 2. Automated weekly client reports. Pulls data, builds tables, emails it I do not touch a thing. 3. Budget pacing and overspend forecasting. Projects end-of-month delivery. Flags blown budgets early. 4. Creative performance breakdowns. Ranks every ad by ROAS, CTR, CPA across both platforms. 5. Cross-channel attribution checks. Flags double-counted conversions ... ... sets cannibalizing each other. 7. Search term mining and negative keyword lists. Finds the negatives I should have added last month. 8. Landing page and tracking QA. Checks every URL and pixel fire in the account. 9. Competitor

... just shipped this in Basedash and my honest first reaction when we scoped the feature was no way, absolutely not. I've spent 6 years building tools that touch production data, and the whole time the rule was that humans click the buttons. But we did it anyway, and the process of getting comfortable with it changed my mind about where the risk actually lives. What got us there: Admins choose which databases the AI can edit at all. Read ... ... only is the default. Every tool and connection has its own permission, so you can let it update Stripe but never touch

... every team deploying AI agents eventually runs into the same shift. At what point do you think an AI workflow needs governance? Comment from Syed Noor(@syed_noor4): The line I've landed on: it's not "when it touches real customer data," it's "the first time it can do something you can't take back." Reading data is recoverable you re-run the query. The moment the agent can write to a system of record, message a customer ... ... governance later you're rebuilding. So, my answer to "at what point": the moment the agent stops only reading and starts doing. That's usually earlier than the demo makes it feel. Curious how you're drawing the read-vs

... invent revenue figures. Not close. Wrong. Confidently wrong. So I split the job. Financial figures come directly from the SEC's XBRL structured data feeds. The AI reads the text and flags red flags. It never touches a number. Every claim links back to the verbatim paragraph in the original EDGAR filing. One click to the source. The result is Sharpread. Forensic SEC filing analysis where the AI reads the filings but never generates the numbers. If you invest ...

... tool   2. ChatGPT / Codex one app now: where I argue about architecture, and the implementation pair for bigger diffs   3. Higgsfield all my image + video generation; product visuals and social content without touching a design tool   4. Wispr Flow I barely touch