ScaleRL isn’t magic. It’s a recipe of small, proven ingredients: – Pipeline-RL async setup
– FP32 logits for stable gradients
– CISPO loss (clipped importance sampling)
– Prompt-level averaging
– Batch-level normalization
– No-Positive-Resampling (drop “too easy” prompts)
PROMPT ENGINEERING
-
ScaleRL: Key Components Explained
By
–
-
GPT-5 Outperforms Sonnet for Idea Brainstorming Tasks
By
–
ChatGPT is the way to go! Specially for bouncing ideas – gpt-5 is way more steerable than sonnet
-
Rubric-Driven Evaluation Improves LLM Alignment and Data Quality
By
–
Rubric-driven evaluation doesn’t just improve human and LLMaJ alignment. It accelerates delivery, reduces rework, and improves the thing we care about most: data quality.
-
LLM-as-Judge Alignment Improves to 94% With Structured Evaluation
By
–
After refining our rubric design process, alignment between human reviewers and LLM-as-a-Judge improved from 37 % → 94 % — proof that structured evaluation enhances consistency.
-
Anthropic Releases Agent Skills for Extended Claude Capabilities
By
–
New on the Anthropic Engineering Blog: Our tips for developers on using Agent Skills, a new way to extend Claude's capabilities with instruction folders, scripts, and resources:
-
Try Out Skills Feature on Claude AI Platform
By
–
Or go try out Skills in claude dot ai: https://t.co/W5GdkvlVKL
— Alex Albert (@alexalbert__) 16 octobre 2025Or go try out Skills in claude dot ai:
-
Building Skills Iteratively with Claude: Best Practices
By
–
If you want to make a Skill, we recommend building iteratively. Run Claude on real tasks, watch where it struggles, then capture what worked. Claude can help you write skills as well with our skill builder skill in claude dot ai.
-
Claude Skills Bundle Files Support Multiple File Types
By
–
These bundled files could contain everything from additional markdown instructions to image assets to code scripts that Claude can execute. Since Skills are just folders, they can accommodate any file type!
-
Skills: Simple Folders for Specializing Claude Without Custom Agents
By
–
At their core, Skills are simple. They're just a folder with a http://
SKILL.md file. The file starts with a name and description, then contains instructions, code, and resources. This simplicity means anyone can now specialize Claude without building custom agents. -
Skills: Loading Specialized Knowledge into General Agents at Runtime
By
–
At a high level, the best analogy I've heard for Skills is something like Neo learning Kung Fu in seconds in the Matrix. We're "loading in" specialized knowledge to our general agents at runtime.