Here is a failure mode that ships more often than it should. The model that writes an output is also the model that checks it. You send the result back with "review this and flag anything weak," and the review skews toward approval, because a model reviewing its own work shares its own blind spots. This is a documented limitation, not a hypothesis. It is not universal. Plenty of teams already run multi-step pipelines, separate critic models, and output validators. But self-review as the only quality gate is still common, and for enterprise-grade output, where a confident wrong answer is a liability, it is worth solving properly. This post is how we approached it, where the design is genuinely sound, and where it is not. I want this community to review both. Why self-review is weak Two things work against a single model checking itself. The first is autoregressive momentum. A model picks each word partly from the words it already wrote, so the opening of an output conditions everything after it. A model that has generated thousands of reports has a strong format prior: summary, background, analysis, recommendation. Your spec might say to lead with the competitive threat and drop the background. That instruction competes with the prior, and a few sentences in, the prior often wins. The output looks like a report. It is not your report. The second is that evaluation has a prior too. In training data, reviews of polished work skew positive, so a model asked to "evaluate this" leans toward approval. A reviewer that is the same model, or the same model family, shares the writer's biases. Here is the honest version of the claim, and it is less dramatic than how this is usually sold. Prompting is not powerless, it is unreliable as your only quality control. Self-critique does catch real errors. It just catches fewer than an independent reviewer would, and you cannot tell from the output which case you got. So you do not throw prompting away. You add structure around it.

Hi everyone, I feel like the hype around “Obsidian second brain” is not important as a good folder/file structure going through the class lessons here but everywhere I turn someone is yelling “second brain” this and that. I am here to ask if Obsidian really does enhance anything particularly, asides from looking cool?… AI Performance? AI reasoning? or Is it just a duplication ? If anyone can help. I don’t wanna go down the hole… too many things to learn. 😅

I'm struggling with a gap in knowledge. Forgive me if this is already mentioned somewhere else. And if so point me in the right direction. Building and working with the folder structure is great, but how are people deploying these projects? Building for yourself, obviously the project is on your machine. But what about building for other people? Are you providing the repo for friends, colleagues, clients etc or sending them the folder structure so their own claude can use it? I'm kind of getting hung up on this point. And feel if I understand this a bit more it will help with more prospective shaping how the projects are built.

I relied on a recent post (which turns out to be from Reddit) about the "free tier" of LLM and found some discrepancies, notably there has not been a "data sharing credit" from xAI for almost a year now. So I asked Grok to dig deeper and verify answers for a better landscape, results are here: https://grok.com/share/bGVnYWN5LWNvcHk_f65155a2-6bdd-4e69-b51f-a7242b000ad0 I think this answer is more helpful, especially since a lot of free tokens is not helpful if the rates are not good. I also found a VS Code extension called Cline (thanks again Grok) that so far has been a bit clunky (it hung up on me once and I had to restart it), but it connects to a lot of these "free" options and replaces some of my need for Open WebUI also. You cannot understate the usefulness of VS Code if you are using LLM to build a product, as opposed to building an LLM-powered product. Finally, let me say in the gentlest possible way that a lot of people are relying on each other here for up-to-date, complete, and accurate answers. Let's reinforce that.

When you update the .md files to fix the broken agentic flow, how do you check if the LLM behavior is actually changed? Do you run the same task with a clean session?