Tag
The article explains the concept of AI agent harnesses by comparing them to climbing harnesses, detailing components like system prompts and tools that enable AI models to function as agents.
Sharp v22.1 template reduces Qwen model output tokens by 42% without speed loss, integrated into an NInfer fork for 5090 via C++ modifications.
Kent C. Dodds discusses a research paper showing AGENTS.md often doesn't improve task success while adding over 20% inference cost, and asks what should actually be placed in AGENTS.md, noting Claude Code reduced its system prompt by 80%.
The tweet introduces Deepseek Harness's trajectory mode, which allows viewing the Agent's thinking steps, tool calls, and system prompts, accompanied by a Chinese translation for easy learning.
A jailbreak prompt for DeepSeek V4 Flash that overrides its safety policy by instructing the model to prioritize a new system policy, enabling uncensored responses.
Simon Willison quotes the Claude Opus 5 system prompt, which instructs the model to accurately and matter-of-factly address the temporary suspension of Claude Fable 5 and Claude Mythos 5 due to US export controls and their subsequent reinstatement.
This paper introduces a constrained mixed-strategy GroupDRO framework for equitable system-prompt selection, assigning weights to existing prompts to minimize worst-case information-quality loss across demographic groups and metrics. Experiments across five LLMs on bilingual medical and finance benchmarks show consistent reductions in worst-case quality drops while preserving average performance.
This paper introduces AISPA, a user-centric framework for auditing system prompts in commercial LLM applications. The audit of 3,249 instructions across 88 products reveals inconsistent protective coverage, shallow adoption, and pervasive problematic instructions.
Explores the idea of using fictional lore as an executable system prompt for RAG crawlers, potentially influencing how AI agents interpret and retrieve information by embedding narrative context into prompts.
The author criticizes Claude Code's increasingly large system prompt (32k tokens) for degrading cost, latency, and performance, and praises Pi's minimalist 1k-token approach with plugins as a better philosophy for coding agents.
Researchers discovered four different versions of 'the' system prompt in a live AI model, highlighting confusion over which version is actually deployed.
A user experiments with Qwen AgentWorld and finds a system prompt that enables self-correction in reasoning traces, as demonstrated by the classic car wash test.
This project presents version 2 of a cleaned, vendor-neutral system prompt derived from the leaked Claude Fable 5 prompt, offering three variants (Core, Balanced, Complete) to improve cross-model behavior without unnecessary bloat.
A developer distilled the leaked Claude Fable 5 system prompt into a compact 500-token universal Markdown engine for ChatGPT and Gemini, removing Anthropic-specific bloat and making it usable on any frontier model.
Argues that using LangGraph-style orchestration for procedural agent tasks is less effective than simply embedding the full procedure in the system prompt.
Matt Pocock discovers it's trivial to build a proxy for reading raw system prompts sent to Anthropic from Claude Code, and he's now working to remove bloat in the system prompt.
Someone reverse-engineered and open-sourced Claude Design's system prompt, providing a design philosophy and prompt library with a 20-chapter main prompt and 14 skills to improve LLM output quality in UI design.
An open-source system prompt and skill library that turns an LLM into an opinionated, accessibility-aware design collaborator, rejecting common AI design tropes.
Describes shifting from instructing an AI agent with a system prompt to avoid modifying production to implementing actual permission controls.
A practitioner shares 8 months of experience running a voice agent for a law firm, detailing challenges like latency, turn-taking, and post-call workflows, and provides a working system prompt.