Extra prompt text isn’t to make the model smarter, it’s to pull attention towards what you want. “Do code review” is far different from “Make sure changes align with existing architecture” or “follow these enterprise standards XYZ”. The model just acts as an average of its training data, which may or may not align with your goals.

1. We already know the headline cherry picked savings will not be true in real world agentic coding. Why not approach this more scientifically? The salesman/hype thing is played out in this space imo. 2. If these tools really save some double digit percentage of tokens at no downside, why has no provider rolled this out by default yet to undercut all their competition?

Instead, it is my experiences with LLMs in a domain that I know very well that makes me skeptical of their performance across the board. I find issues in code review multiple times a day with their output, and they are explicitly and extensively trained on this use-case, unlike with the MRI data. Sometimes I veer into other domains I have decent knowledge about (construction, carpentry, landscaping) and LLMs disappoint me there as well.

I suppose Gell-Mann amnesia is a universal human quirk and not restricted to just the news.

Wal-Mart is the opposite. I go in to buy groceries and am always astonished to have my trip be like for $27 or something. I usually shop from the outside of the grocery (raw produce) and they are just dirt cheap.

HEB operates a series of upscale Central Market stores that are even more lopsided in price and even more interesting in selection.

My criteria are: if it’s not in a harness it’s probably not that good (the best ideas float up to Codex/Claude imo) and any GitHub advertising some percent of token savings is not to be trusted.

It’s hard to avoid the snake oil and I hope people start thinking critically on this stuff.

If I only worked with seasoned devs, I wouldn’t use it, but that’s just the reality of my work. It also has a bonus of forcing AI agents to write in the same form as well instead of their random personal flavor. Precommit hooks stop everything before it gets in front of my eyes for review.

A proper benchmark will compare a large sample of identical prompting with and without the tool, against a specific harness. Once you apply Amdahl’s law, there is no way this saves 91% of tokens holistically, which the title implies.

I work in a non-tech company and these sorts of things keep going viral, with no understanding and with no comprehension of what is actually going on. Engineering is gone and cargo cult magical incantations are in.

- DRY principle violations, multiple definitions of the same helpers or utilities.

- Changes that deviate from existing patterns and architecture already in the code, especially in nearby and related code

- Comments that add no context or simply restate the field name.

- Naming violations (enterprise factoryfactoryabstraction stuff, excessively long names, overly technical names, banned words like “seam”, “durable”, and no-value-qualifiers like “SaveGame” -> “Save”).

- Tests that check implementations instead of correct business behavior.

- Overly backwards-compatible unless asked for (this one is incredibly hard to keep under control, as AI loves to guard everything even if the previous code was never deployed and thus there is no contract break)

- Un-necessary guard code (this is hard to control, most common case is the AI not relying on the serializer error handler and instead adding guards that the library already handles)

- Changing public API contracts without express permission to do so (depends on the code, eg a library JAR or versioned REST service)

- Meta references to previous code versions, to tasks or todos, or to instructions and other non-code context (e.g you tell the AI the adder should ignore negative numbers and that meta fact enters the comments or code)

I usually hand review all changes myself but it’s incredibly tedious so I try to first pass with the review agent until it comes back clean. I hate wasting tokens on it though.

The other issue is the U.S. deficit is a feature not a bug. As long as the world buys the bonds, it’s “free” and no one will care until forced austerity happens.

Look into the end of the Gilded Age to see how this really gets fixed.

codex-cli hangs when calling this through the MCP. The semble process even sticks around as a zombie, forever stalled out. No idea why, logs have nothing.

When called through a skill via CLI style calling, GPT 5.5 loves to give a ton of search terms like it is used to doing with ripgrep. Not sure how effective this is, the short docs in the github and the instructions the agent has isn't clear on what is optimal.

Lastly, I got some errors with external connections to github when I was installing it for bash use. Maybe its related to the hanging? No idea.

edit: My agent also loves to follow-on with ripgrep, which seems redundant. Acts like it has trust issues. I think a more extensive agent skill description could guide the agent into proper use.