Boris from Anthropic: delete all skills and the claude.md file Delete the noise. Keep the signal. Modern models do not need endless prompts, bloated skills, or rigid step-by-step control. What Boris Cherny makes clear is simple: Test without the scaffolding. Observe where the model truly fails. Add back only what proves its value. That is exactly what this audit prompt is for: - find outdated instructions; - remove redundancy; - cut micromanagement; - preserve essential context; - keep guardrails; - strengthen verification; - rebuild only what is necessary. I spent a lot of time studying this, collecting information, and testing my own system. That led me to create an Ablation Audit Prompt to analyze my CLAUDE.md, skills, and overall setup. And I discovered a lot. I was able to significantly reduce my skills by disabling always-on search and turning it into something that only runs when explicitly called by command. I moved instructions and context into project-level files so they are loaded only where they are actually needed. I also optimized skill descriptions, removed redundant instructions, and simplified the system. The result was better than I expected: Fewer tokens. Less noise. Better reasoning. More autonomy. In my case, this optimization became even more useful than relying on /doctor, because it helped me understand exactly where complexity was coming from and what could be removed, moved, or loaded only on demand. One of the most important lessons came from comparing model generations. Everything I had done to improve performance with skills and Markdown files worked extremely well with Opus 4.8. But with Opus 5, some of the same instructions started making the system worse. Once I removed or simplified them, the model improved again. That finally made the pattern clear: What helps a weaker model can limit a stronger one. Think of it like giving instructions to a child, a teenager, and an adult. A child needs detailed steps.

BMAD x Superpowers x ECC Separates BMAD, Superpowers, and ECC in a very simple way: BMAD = think and structure the project. Superpowers = execute the engineering work. ECC = expand the agent environment with additional capabilities. For small projects and everyday development tasks, such as features, bugs, refactoring, testing, and debugging, the recommended setup is: Claude Code + Superpowers For large or complex projects, the recommended flow is: BMAD → Superpowers BMAD handles the strategic side: discovery, project definition, PRDs, UX, architecture, epics, stories, and planning. Then Superpowers handles execution: understanding the task, planning, implementation, testing, debugging, review, and verification. ECC does not replace either of them. Instead, it works as an additional infrastructure layer that can provide capabilities such as: persistent memory security hooks automations MCP integrations specialized agents research and data collection specialized reviewers language- and framework-specific skills shared standards across projects The decision logic can be summarized like this: Not sure what to build? → BMAD Already know what to build and need to implement it properly? → Superpowers Need an extra operational capability for the agent or environment? → ECC The recommended configuration is: Claude Code + Superpowers → the default for most development work. BMAD → Superpowers → for larger or more complex projects. ECC → added selectively when there is a concrete need. The key idea is not to install or use the entire ECC stack by default. Treat ECC as a modular catalog of capabilities. Add memory, security, hooks, specialized agents, or other components only when they solve a real problem. In one sentence: BMAD thinks, Superpowers builds, and ECC enhances.

OPENAI api 90% price reduction. Percentage drop - Luna: from $1.00 → $0.10 input and $6.00 → $0.60 output = 90% price reduction. - Terra: from $2.50 → $1.00 input and $15.00 → $6.00 output = 60% price reduction. Summary: Luna fell 90% and Terra fell 60% from the original price to the current Batch price. How many times cheaper - Luna: 10× cheaper - Terra: 2.5× cheaper In other words, with Batch today, Luna costs just 1/10 of its original price, while Terra costs 40% of its original price.

### MCP: What Changed and What’s the Advantage? **Before: MCP needed to maintain a session and “remember” each user between calls. This made the infrastructure more complex and could require persistent sessions, Redis, sticky sessions, and tighter control over which server handled each user. **Now: each request can be independent. The client sends the necessary context with every call and, when a process needs to continue, it can use a secure identifier — similar to the “restaurant tab” analogy used in the original explanation. In practice: ```text BEFORE User → session → specific server → keeps state NOW User + context → any server → response ``` ### The Main Advantage For developers: * less complexity; * simpler deployment; * less dependency on session infrastructure; * better use of serverless and edge environments; * easier scaling; * simpler retries and load balancing. For users, almost nothing changes in the interface. The difference happens behind the scenes: the system can become more scalable, resilient, and cheaper to operate. In one sentence: > MCP is moving away from a conversation tied to a specific server and becoming more like a modern HTTP API: simple, independent, and easy to scale. And that matters because it makes it much easier to turn existing APIs, systems, and tools into capabilities that AI agents can use. https://modelcontextprotocol.io/specification/2026-07-28/changelog

Short Summary of AI Loops AI loops are cycles in which an artificial intelligence system performs a task, checks the result, corrects failures, and repeats the process until it reaches a goal or a defined limit. There are three main types: - Turn loop: the AI reviews its own work before responding. - Goal loop: it keeps working until verifiable criteria are met. - Time loop: it repeats a task at scheduled times or intervals. In Claude Code, /goal keeps the AI working until a condition is proven, while /loop repeats a task at a time interval. Stop Hooks can control whether work continues after each turn. Auto mode makes tool use easier, but it does not create a complete loop by itself. A reliable loop needs a clear objective, success criteria, evidence, time or attempt limits, and a stop condition. Loops can consume many tokens when they involve large contexts, many retries, multiple agents, or repeated checks. The main idea is: Do not simply ask the AI to keep going. Define the goal, how it will be verified, when corrections should happen, and when the process must stop. A well-designed loop increases autonomy. A poorly designed loop increases errors, rework, and cost.