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.

Running Claude Code on Opus is powerful. But most turns in a session are mechanical, reading files, writing boilerplate, simple edits. You're paying top-tier rates for work that Haiku could handle in its sleep. Anthropic just released the Advisor Tool: A server-side pattern where a cheap executor model consults Opus mid-generation for strategic guidance. But it's API-only. Can't use it inside Claude Code. So I rebuilt the pattern inside Claude Code using what already exists. The setup: Opus stays in the orchestrator seat. It plans. It makes architecture decisions. It reviews output. It never touches files directly unless the task genuinely needs Opus-level judgment. Everything else gets dispatched to cheaper models: - Agent({ model: "haiku" }) — Claude subagents with full file access for simple edits - Agent({ model: "sonnet" }) — for multi-file changes that need moderate reasoning - A CLI tool (ask.py) that routes to Gemini, Kimi, MiniMax, or local Gemma via Ollama — for code generation, research, video analysis, anything where you just need text back The routing logic: Does the task need file access? → Agent tool (haiku/sonnet) Is the code complex? → sonnet or gemini Is the code simple? → haiku, kimi, or minimax (cheaper) Need video analysis? → gemini --video (native, no frame extraction) Need parallel research? → kimi-swarm (spawns up to 100 sub-agents) Want zero API cost? → gemma running locally via Ollama When a subagent hits something it can't handle, it reports back NEEDS_GUIDANCE. Opus thinks it through and re-dispatches with better context. That's the advisor pattern, strategic guidance exactly when needed, cheap execution everywhere else. Cost impact: A 10-task session where 8 tasks go to Haiku and 2 to Sonnet — your Opus tokens are only spent on planning and review. Maybe 20% of the total token volume. The other 80% runs at Haiku rates. With local models mixed in, it drops further. Open-sourced the whole thing:

Has anyone else noticed Claude was actually really unintelligent this week after mythos preview was announced? thank god for Clief’s Folder Architecture, all I had to do was hard gate into global Claude config and Claude become intelligent enough to use again PSA: don’t use open sourced models for coding + reasoning, no matter how good they are on benchmarks if you plug it into Claude code it will never be as good as Claude code, just take a couple days break Open sourced models can do task execution well but other than that honestly don’t waste your budget

Is everyone caught up with all the new claude code features? so many new features i don't know what new capabilities i should start learning 🥲

Has anyone tried to apply the folder structure learned here into their obsidian vault? Clief folder architecture really helped with retrieval and how knowledge is linked, this new meta really documents my current understanding on various subjects very well, in claude code sessions i don't have to explain as much context as before because claude code "knows" what i mean even better because of obsidian vault, but i realized that there needs to be a naming convention that works with clief folder architecture to make it truly optimized Anyone have any suggestions as to how to set up the naming convention?