Claude Opus 5.0's Incoherent Output Sparks User Backlash

This drives me insane! To combat it, I maintain a list of banned words and phrases. Claude mostly follows this (but sometimes ignores it).

blast radius, land, landed, lands, spine, earned its keep, grammar, spike, cutover, bake, seams, honest, honestly, honesty, long pole, long poles, register, grain, dissolve, floor, ladder, dear, seal, sealed, in anger, resent, amazing, incredible, perfect, sprint, epic, story points, stand-up, retro, grooming, robust, comprehensive, rigorous, surgical, elegant, systematic, dive, deep-dive, delve, unpack, leverage, streamline, surface, it's worth noting, to be clear, importantly, that said, the moment, in one breath, the thing itself, here's the thing, not just X but Y, not X it's Y, em-dashes

Yeah I basically have to delete or rewrite any code comments Opus puts in, they are super wordy and circular. Plus, and this is what bothers me most: it refers to issues/explanations that came up in my chat with it, but which are completely contextless as a code comment. Stuff like 'this addresses edge case X', which is true but which you wouldn‘t ever put into a comment because it‘s so minor, and if you did bring it up you normally mention five other similar edge cases.

I suppose what I’m surprised by most is the degree to which the text bloviates. Rarely is it encoding a dense concept. Instead it’s a simple concept expressed with great complexity. I wonder what tuning resulted in this because the output programs it writes are superior to prior iterations for my use-case.

It’s just an amusing degree of bombast as well. The pre-emptive hedging makes sense: despite the insight into J space etc., the models still do the majority of their thinking in generated tokens so it is forced to write “this no longer does an O(n^2) read over all rows” in a comment in brand new code. It’s a substitute for working memory. “It’s easy to be accidentally quadratic here, so I’ve done it this way specifically to avoid that” becomes temporally labeled into “this no longer” because of the order of operations “write quadratic, user prompt to linear, write linear” but it remains as a comment to its amnesiac future self which has poor Chesterton-Fence-familiarity.

Despite my annoyance with Claude’s writing style, my friends do tease me with examples like this that it comes up with: “to be honest, it sounds like you”. Thanks, guys, well played. Simple concepts expressed complicatedly.

Basically every decision or summary I now ask "Help me understand better, please explain in plain-language with examples."

Works with 5.6 sol also, when you're deep in the weeds. I rationalize this as the models attempting to compress as much into the fewest tokens, though the choice of words often doesn't make sense to me, going back to read the original after, its often there. It definitely feels like a different sort of 'Machine Language' though xD

I've had a lot of success combating this by adding "All summaries need to adhere to ASD-STE100 Simplified Technical English standards" [0] which I discovered from another HN thread [1]

[0] https://www.asd-ste100.org/

[1] https://news.ycombinator.com/item?id=49114639

Opus 5 feels like a downgrade from Opus 4.8 overall. It, along with Fable, really has a problem following instructions and staying in scope, and their prose keeps growing, both in explaining what it did and in writing multiline code comments (some comments read like a changelog, e.g. `// sky is blue (changed from red on 2026-01-01 per TCK-234 by @Foo)`).

Every time I ask it to do something, it does 80% of the job, goes off on "side quests" beyond the scope, and then leaves something out of the core ask (and when you tell it to finish, it does the same thing again).

The only advantage of Opus 5 over 4.8 is the better cutoff date for working with 3rd-party tools, though both do a very bad job of "this tool is constantly updated, I should look for the latest version first".

Yeah it's really, really bad. It's borderline unreadable. Claude seems to choose the most clever word choice it possibly can instead of the simplest. It's like the perfect inversion of how writing should be. Why say "load-bearing" when you mean "important"? Also, 50% of the stuff can't be load-bearing.

Another problem I didn't see mentioned in the thread is the conversation and reasoning leaking into text. It's a big problem for code comments. A comment will include multiple tirades about what we decided NOT to do.

Here's my list of mitigations:

- Keep sessions short

- Remind Claude of writing style. It will only last maybe 1-2 prompts as the thread notes, but if the session is short it helps

- Plan and implementation should be separate sessions to avoid the conversation leaking. My workflow: plan and brainstorm, scaffold APIs and tests, then have Claude write a seed prompt for the next session. I will also iterate over the seed prompt because it has the same text issues.

- Instruct Claude to compact comments often and have a rubric. Describe WHY, never WHAT. Comments should prioritize simple language. Etc.