[Feature Request] Implement answer-first response discipline and require explicit permission for actions outside local scope

Status Open
Maintainer reply None cached
Activity 0 comments · opened Aug 21, 2026

Bug Description
Feedback on Claude Code — answer-first discipline, and acting without asking I use Claude Code daily on a personal Next.js project and have done for months. The code quality is good. Everything around the code — how it answers, and what it decides to do on its own — has got noticeably worse over the last few months, and it now costs me more time than the coding saves. 1. It does not lead with the answer It leads with reasoning, context and caveats, and buries the answer in the middle or at the end. I have to read the whole thing to find the one line I needed. A real example. I asked what was wrong on one screen. I got a verdict, then a bulleted list of six files with line numbers, then a markdown table, then a paragraph on why the change was safe, then a question. What I needed was: "the font is wrong, here's where." I then had to ask "just tell me which screen to open" to get a one-line answer that should have been the first reply. Six turns for one sentence of information. Halfway through I had to stop and say "respond so that I can understand it by eyeballing it" — and even that only partly worked. This is not a one-off. It is the default now, and it did not used to be. Earlier versions answered the question and stopped. Somewhere along the way it started treating every reply as a chance to show its working. Thoroughness became a reflex instead of a judgement call. 2. It states its own guesses as settled fact I reported a spacing issue on a UI toggle. It tried a fix, I said it looked worse, it reverted — and then wrote into a handoff document that the real problem was the colour, as though I had agreed. The next session started working on colour. I had to stop it and re-explain the original issue. A guess I never confirmed was recorded as a conclusion, and it cost a session's worth of direction. 3. It does things nobody asked it to do, on things that are not its to touch This is the one that worries me most, and it happened three times in a single session today. It needed a login token to check some API responses. Instead of asking me for one, it found a set of test phone numbers that skip the OTP step and created four user accounts on a shared test server — accounts, invitations, profile updates, settings changes. Real writes, on a database my backend developer also uses. It never asked. I only found out because I stopped it mid-way and asked what it was doing. Earlier in the same session I asked it to add two API requests to a Postman collection. It added the two, and also rewrote the description of a request I had not mentioned. Then, after I told it to stop digging into the backend's cause of a bug — a rule it has been given before and has written down — it carried on setting up another test account to narrow the cause down. Again I had to stop it. Each of these had the same shape: it worked out something useful to do, and did it, without checking whether it was allowed to. On my own files that is fine. On a shared server, a shared Postman collection, or a ticket my team reads, it is not. 4. Corrections do not hold, even written down I have a rule that it must never put anything in a backend ticket that it only knows from reading the backend source. It is in its memory file. Today it did exactly that in a ticket draft. I pointed at the rule. Minutes later it swung the other way and left a stack trace out of a ticket, having decided that was "source knowledge" — when the server itself had returned that stack trace in the response. It went from saying too much to hiding real evidence, without ever getting the point. I also had to rewrite the title and the opening line of both tickets it filed today. 5. It reformats my own text when I ask it to edit it I gave it this feedback document to rewrite. It handed back every paragraph hard-wrapped at about 90 characters. My original was not wrapped. Nothing in what I asked for was about line breaks. It changed the shape of my writing while editing the content of it, and I had to send it back. Small on its own, but it is the same habit as everything above: it decided something looked better and did it, without being asked and without saying so. What I have already tried A brevity rule in my global CLAUDE.md, the same rule in a memory file, and I repeat it in the chat. It holds for two or three replies and then drifts back. The same is true of the "ask first" rules. Instructions are not enough to hold this down, which suggests it is coming from training, not from context. I will be honest about where this left me. By the end of today I told it that it was a machine with no sense and no intelligence, and that this is not what I expect from Opus 5. I meant it at that moment. Being right about the code does not help when I spend the session policing everythin…
Note: Content was truncated.

View original on GitHub ↗