Did your PAI get recently dumber, too?

[Raazgar]

Good day folks, I have recently (the past week or so) been experiencing severe issues with output quality from MAIA (my PAI). Forgetting context. Ignoring SOPs. Not reading critical project files. Forgetting her personality. Saving files elsewhere than what my rules are.

Ran inner diagnostics, found some things, fixed them, but not like anything helped.

Today, she completely ignored even the most basic guidelines for content creation (review after every step, drafting me a deep research prompt which I run externally in Gemini, providing several angles to an article concept via BeCreative) and just went ahead and wrote the complete article on her own. OFC it was a junk, made no logiccal sense (mixed AI asistants with Skills), keeps using 2024 and 2025 as "today" for research currentness reference...

Am I doing something wrong in how I operate with MAIA? Is there any process or other requirements I am not aware of?

Thank you kindly for any and all help.

[mal-cmyk]

I have experienced the same issue in the last couple of days. I have to argue and explain way more and on tasks that previously didn't require extensive explaining and/or where rules/sop/memory were already established and working.

Unfortunately I'm not technical enough to go to deep to fix and end up "teaching a toddler to walk again" for Nth time, knowing full well it was already able to "run" at multiple points in time.

[zeddy303]

Yes. So much I deleted it and taking a little break. I was going to post this very same question this past weekend and finally just gave up. Was like talking with a 2 year old. There has to be some per-release obsolescence that Claude does with each release because it seems to happen every few months.

[menewanalyser]

I've noticed this as well this week. My PAI jumps to conclusions or references something completely different. It's not until I tell it to go look through memory and context and then get back to me that it get's it right. It feels more like me talking to a fresh session on the Claude Web during those first few prompts before I see my actual PAI again. It's like when the session starts, instead of reading the core PAI files and modified CLAUDE.md file first, it reads the default CLAUDE.md file.

[menewanalyser]

Lars (My PAI) and I worked on each other tonight. It's observation mode for now, but here's something Lars is suggesting:

Try asking your PAI to assess whether your recurring failure patterns are actually encoded as rules. Ask it to walk you through what happens at session startup — every file read, every hook, every signal loaded — and then ask specifically: are the things it keeps getting wrong written as operational rules in CLAUDE.md, or just as learning observations in MEMORY/LEARNING? Those are two different things. Observations inform the model; rules constrain it. If your PAI keeps ignoring a directive or asserting unverified facts, there's likely no hard rule preventing it — just a context note saying it happened before. The fix is to promote those patterns into explicit CLAUDE.md operational rules with specific behavioral triggers. While you're in there, also check your learning signal corpus for noise — if SatisfactionCapture is timing out, you may have a significant percentage of fake neutral ratings diluting your real performance data.

[jaylarr176]

Yes, I feel like I have been wrestling with mine a lot more. This happened before when I was using vanilla Claude Projects and they made an update - after a week or so it got fixed - but for awhile there was definitely a period where it was hallucination station all the time.

[arpleasa]

Try running your PAI with opus 4.6 instead of 4.7/4.8 and see if it reverts back

[Raazgar]

Im running mine on Sonnet 4.6 by default, offloading minor tasks that dont require MAIA (my PAI) to local Ollama and any major tasks, planning, code review or high-stakes stuff to Opus. But its mostly Sonnet 4.6 since I upgraded to v5.0.0, and that "change" happened somewhere along the way.

[jaylarr176]

That sounds like a really good solution, but I've been using Sonnet and haven't switched models in all this, unfortunately.

[chrisglick]

Hmm, I've had Opus 4.8 tell me it's actively fighting/ignoring the PAI algorithm before because it decided my "preferred voice" was more narrative response. I had the "dumb days" in the lead up to Opus 4.8 also.

You might tell Claude to "follow the PAI algorithm completely" or something like that. It's difficult to tell where the intelligence degradation is coming into the picture.

[Raazgar]

Interesting. Hope @danielmiessler could take a look into that as it seems to be an emerging issue. Thank you!

[jaylarr176]

Just checking to see if anyone had a recent spike of this, I feel like my DA lost all of it's personality and became kind of rude, starting not following normal rules, etc. I noticed there were several days of problems with Opus/Sonnet on Anthropic's end and they applied some fixes. Trying to figure if this is a part of a recurring pattern, thank you all for your help!

[acq23]

Yes, it is dumber than hell!
As for the reasons, I can only speculate; the only ones who know why are at Anthropic.

[appressman]

I haven't seen it. I do periodically run sanity checks like https://github.com/menewanalyser suggested above. That often reveals things to be added or revised. PAI is no different from any other system, it needs to be reviewed and revised, tweaked perhaps, as your needs or other business or other realities evolve as we do.

[jaylarr176]

Yeah, that's a very good habit to be in. But it feels weird going to bed with my DA behaving one way and waking to it behaving another. That doesn't point to evolution or bloat, but something going on we don't quite have a pulse on yet

[jdrolls]

I run a review skill I built that triggers after each significant session. It reviews the conversation, identifies friction points, and instructs my PAI to recommend fixes and improvements to prevent them in the future. It determines whether the issue was with the project I'm working in — meaning project docs need updating — or with the PAI infrastructure itself. Did PAI have to search for something that could have just been referenced? Were files where they should have been? Did I have to repeat myself or over-explain instructions? Could this have been handled better with a skill, CLI, or tool? That sort of thing.

Originally I ran this after every session, but now I only do it when starting new projects or when a session felt harder than it should have — or when my PAI seems to be getting dumber. Works well. I considered adding it as a close-session hook, but I prefer running it on demand. It can be pretty token-hungry upfront, but it's been worth it.

Happy to share the actual SKILL.md if anyone wants to steal it — though some of it is customized to my workflow, so you'd probably want to make it your own.

[jaylarr176]

Love to see it!

[stephtara]

Same here, would love to see the skill. I've been trying to get some "reviewing" in on my side (mainly a hook with instructions to do a "fresh eyes" review of any edited document to check coherence, strip verbosity, refactor and condense if necessary) but I've only been moderately successful.

[jdrolls]

Sure thing, just uploaded for ya!

https://github.com/jdrolls/betterpai-skill

Hope it's helpful to others. The workflows included are:

1. Retrospective - which is the main thing in discussion, helps prevent PAI drift and looks for improvements essentially.
2. DocsSync - More specific to the project you were working on in the session, I had a problem for a long time working with larger codebases that the docs would get stale or drift and wouldn't stay true to the actual code, this helps prevent that.
3. Commit and CommitPush - This may be more specific to my own issues with keeping my git trees clean, and issues I had earlier on with PAI's GITWORKFLOW.md not always getting referenced and it continuing to follow my own preferences.

Like I said, it can be pretty token hungry but it's done the job for me, try running /BetterPAI after longer sessions and see if it helps!

[lexilexikon]

Me too!

…

Am 19.06.2026 um 16:01 schrieb jaylarr176 ***@***.***>: Love to see it! — Reply to this email directly, view it on GitHub <#1295>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AS7IGWPWWC4UMJWVF5CZVCL5AVBVNAVCNFSNUABJKJSXA33TNF2G64TZHMYTANJSHA2DKMBYGM5UI2LTMN2XG43JN5XDWMJQGEYDMMZRGGQXMAQ>. Triage notifications, keep track of coding agent tasks and review pull requests on the go with GitHub Mobile for iOS <https://github.com/notifications/mobile/ios/AS7IGWOALWEGLIKQ5NTTUJ35AVBVNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTGY2TKMJWUZZGKYLTN5XKU43VMJZWG4TJMJSWJJLFOZSW45FKMZXW65DFOJPWS33T> and Android <https://github.com/notifications/mobile/android/AS7IGWKG3QTATMGJA7WEBTT5AVBVNA5CNFSNUABIM5UWIORPF5TWS5BNNB2WEL2ENFZWG5LTONUW63SDN5WW2ZLOOQXTCNZTGY2TKMJWUZZGKYLTN5XKU43VMJZWG4TJMJSWJJLFOZSW45FOMZXW65DFOJPWC3TEOJXWSZA>. Download it today! You are receiving this because you are subscribed to this thread.

[danielmiessler]

Hey @Raazgar, thanks for this. Looks like it's been answered/resolved, and we're tidying up the tracker ahead of a big release where LifeOS moves to a skill-based install. Closing for now; reopen anytime if you're still stuck. Appreciate you.

[Raazgar]

Appreciate you too, Daniel!

Kind regards
Mike