Claude Agent Skills: A First Principles Deep Dive

kept by eddie nov 25, 25

Claude's Agent Skills use a prompt-based meta-tool architecture that injects specialized instructions and context modifications to guide complex workflows without executing code.

Transitions.dev — Essential transitions for web apps
A curated collection of essential UI transitions for web apps, designed to be easily copied and pasted or integrated via the Skill tool.
also on Skill, #developer-tools
rl-wrong-about-rewards.md
The standard reinforcement learning formalization errs by placing reward in the environment; instead, reward should be part of the agent, enabling goal-driven behavior.
also on #agent-design
Changes in the system prompt between Claude Opus 4.6 and 4.7
Opus 4.7's system prompt adds child safety tags, a tool-search mechanism, anti-verbosity instructions, and drops the Trump presidency clarification from the 4.6 prompt.
also on #prompt-engineering, Claude, Anthropic, #developer-tools
Caveman: Why use many token when few token do trick
Caveman is an open-source Claude skill that forces LLMs to reply in terse, ungrammatical speech, cutting output tokens by up to 65% while preserving answer quality.
also on #prompt-engineering, Claude, #ai-coding, #developer-tools