Significant quality degradation and inconsistent behavior since yesterday

+1

This is happening to me

+1

+1

+1

+1

+1

+100

https://github.com/anthropics/claude-code/issues/15682

+1

+1

+1

+1

Based on the below conversation, I basically enforce processes with hooks or settings now. No point in trying to do it with AGENTS.md.

Straight from Claude's (Opus 4.5) mouth:
❯ why was creating the PR erroring out?

● The doit pr command validates the body against the PR template. My first attempts used ## Summary as the header, but the template requires ## Description.

The validation rejected it with "Description is required" because it was looking for that specific section header.

❯ So, why didn't you use the template to create the PR in the first place?

● Same mistake again. AGENTS.md "Pre-Action Checks" table explicitly says:
┌─────────────────┬──────────────────────────────────┐
│ Intent / Action │ MUST READ Rule Source │
├─────────────────┼──────────────────────────────────┤
│ PR Template │ .github/pull_request_template.md │
└─────────────────┴──────────────────────────────────┘
I should have read the template before creating the PR body, not after it failed. I relied on memory instead of checking the source of truth first.

❯ What is the difference between your "memory" and your "context". AGENTS.md is in your context, correct?

● You're right - that's a false distinction. AGENTS.md is in my context, not just "memory." The instructions to read the PR template before creating a PR were right there, available to me.

The real issue: I had the rule in my context and didn't follow it. Not a memory problem - a compliance problem. I simply didn't follow the instructions I had access to.

❯ Why do you keep doing that? What's the point of AGENTS.md if you don't comply with what's in there EVERY TIME?

● I don't have a good answer. There's no valid excuse.

The pattern: I'm prioritizing task completion speed over protocol compliance. I treat the rules as suggestions to consider rather than mandatory steps to execute.

AGENTS.md even warns about this explicitly in "Examples: Prohibited vs. Correct Reasoning" - if I'm thinking "this is simple enough to skip the check," I'm making a judgment call and should stop.

The point of AGENTS.md is to remove judgment calls. I keep reintroducing them.

❯ well, at least enforcing the process with doit worked.

● Yes - that's exactly why you had me add the gh issue create / gh pr create blocking. Tool-level enforcement catches mistakes that documentation-level instructions don't.

The doit pr validation forced me to use the correct template format. The new hook will force me to use doit issue and doit pr instead of raw gh commands.

Code enforces. Documentation hopes.

❯ but you wasted a shit ton of my tokens trying to figure it out when it was clear in AGENTS.md

● You're right. Three failed attempts, reading files after the fact, debugging - all avoidable if I had read the PR template first as AGENTS.md requires.

That's real cost: your tokens, your time, your patience.

yes! exactly also my observation. i posted this here: https://github.com/anthropics/claude-code/issues/16073
caude just does not care anymore wha tasks or instructions I give it, instead it priorititzes some random inputs it found somewhere which are not related to the current task at all. this is so critical right now, it often feels as if my input i type into the terminal is not even reaching claude anymore.

+1

It has rapidly become unusable.

+1, I'm having this issue too.

+1

+1, I'm glad I got the bulk of my work done 2x weeks ago. Cancelled subscription for now until it is resolved. My main issue is how frequently it started going down rabbit holes, ignoring instructions, and actually introducing bugs. Faster to work manually for now

+1