1. What you do. No company names.
2. One thing you got an AI tool to do recently that you could not have done before, or would not have had time for.
3. One thing that still fails every time you try it.

That third one matters. It is usually the start of a real question, and it tells the people building material here what to work on next.

One rule for this thread: if you can help with someone's number 3, do not answer it here. Start a topic in Questions and help and mention them. Intros stay intros.

Put the question in the title. "Agent edits files I did not ask it to touch" gets opened. "Need help" does not.

One problem per topic. Two problems in one post get one answer, usually to the easier one.

Say three things: what you tried, what you expected, what happened. Paste the exact error text. A description of an error ("it fails with some permission thing") is not the error.

Name the tool and the version. Claude Code, Cowork, whichever it is, and the version you are on. Say what kind of project it is: language, size, roughly how old. No employer names, no client code.

Search first. The problem may already have a thread. If it does and the answer does not fit your case, reply there instead of opening a copy.

When it is solved, say what fixed it. The next person with the same problem finds your thread. A thread that ends in silence helps nobody, including you in six months.

What is where

• Questions and help. You are stuck getting an AI tool to do something on your own project. Ask.
• Show your work. Something you built or fixed with AI, and how it went.
• The Workshop. One forum per course or book, for everyone working through it.
• What do you want to learn next? The course or guide you wish existed.
• Announcements. News from the people running this place.

How joining works

A new account can read everything and post in one place: the Introduce yourself topic in this section. One reply there and the rest of the forum opens up. It takes a minute, and it means everyone here said hello before they posted anywhere else.

The rules

There are five, and they are enforced.

1. Stay on topic. This is about doing real work with AI. No off-topic threads, no general chat.
2. No promotion. No links to your product, your service, your affiliate page or your profile. Links in posts carry nofollow, and moderators remove promotional ones. If you built a tool that solves the problem in a thread, describe what it does in the thread.
3. No employer names, no client code. Describe the problem and change the names.
4. Be direct, be decent. Disagree with the idea, not the person.
5. Say what fixed it. When your question is solved, come back and say what worked.

Breaking rule 1 or 2 gets the post removed. Doing it twice gets the account removed.

I request a change, I get back a diff somewhere between 300 and 800 lines, and my review degrades into skimming for anything that looks alarming. Twice now I have approved something that was obviously wrong in hindsight, and CI caught it, which means the review contributed nothing except a delay.

Reviewing a colleague's 600 line change would also be bad, but at least I could ask them why they did it that way. Here there is no why beyond what I asked for.

I do not want the answer to be discipline. Discipline is the thing I ran out of. What is the process?

We were asked for a summary of break clauses across a portfolio. 60 leases, abstracts in a mix of Word documents and scanned PDFs, produced by three different firms over about fifteen years. Normally this is two people for a week.

What I did: I wrote out the eight fields I wanted and, for each one, what a valid value looks like and what to do when it is missing. That brief took most of an afternoon and it is the only reason the rest worked. Then it went through the documents in batches of ten and produced a table with a source quote next to every single value.

The source quote was the whole trick. Checking a row took me forty seconds instead of going and finding the clause myself.

Four rows out of 60 were wrong, and all four were the same mistake. Where a lease had been varied by a later side letter, it took the original break date and ignored the variation. Once I knew that, I checked all 60 for side letters directly.

Two days, one person, and I trust this table more than the handmade ones, because every value in it is quoted.

Anything inside one package is fine. Anything that crosses packages gets me an answer that reads well and is wrong, because it never opened the file where the behaviour actually lives. I have watched it describe a validation rule confidently, based on nothing but the class name.

What I have tried. Feeding it the file tree, too big to be useful. Pointing it at specific files, which works but requires me to already know the answer. Asking it to search first, which helps until it stops after two or three hits and decides it has enough.

What are people doing on genuinely large old systems? I am not asking for it to understand the whole thing. I want it to reliably notice when it has not read enough to answer.

I have a stock movement export, about 4000 rows a month, out of a system that writes dates three different ways depending on which depot keyed them in. I want a monthly summary by product line with returns split out.

When I ask for that I get something that looks right and is not. Last time it quietly dropped every row where the date would not parse, about 300 of them, and handed me a clean looking total that was 7 percent light. It did not mention doing it.

What do people put in the request to stop that happening? I have a feeling I am asking the wrong way rather than asking for the wrong thing.

The extra changes were not wrong exactly. They were just not asked for, and on a client site every extra line is something I have to explain if it breaks next month.

What I have tried. Telling it in the prompt to only touch the one file, which works maybe four times out of five. Listing the files up front and having it confirm before writing, which works better but is slow and I forget to do it when I am in a hurry.

Is there anything more reliable than remembering to ask nicely? Latest version, macOS.