Skip to content

research-background.md

Latest commit

 

History

History
46 lines (37 loc) · 4.84 KB

research-background.md

File metadata and controls

46 lines (37 loc) · 4.84 KB
Raw

Research Background

This project treats skills as lifecycle-managed agent infrastructure: discover repeated workflows, audit and personalize existing skills for local use, and generalize personal skills for publication.

Design Claims

Claim Support Applied In
Skills should be portable, progressively disclosed artifacts rather than one huge prompt. Agent Skills for LLMs survey describes portable skill definitions, progressive context loading, and the relationship between skills and MCP. skill-generalizer, static audit checks, reference/script split.
Agents can improve by externalizing experience into skills or memory without changing model weights. Memento-Skills frames reusable structured markdown/code skills as evolving external memory; Reflexion stores verbal feedback in episodic memory; ExpeL extracts knowledge from experience. skill-miner, archive/session mining, audit evidence.
A growing skill library needs retrieval/routing quality, not just more instructions. Memento-Skills uses a skill router; RAG-MCP shows retrieval can reduce prompt bloat and improve tool selection accuracy; tool-description quality work studies how descriptions affect tool use. skill-personalizer trigger fit, undertrigger checks, broad-cluster vs workflow-candidate split.
Long instructions should keep critical trigger and safety details near the front. Lost in the Middle shows long-context models use information best near the beginning or end, with degradation for middle-position evidence. Frontmatter/description checks, concise SKILL.md, references for details.
Learned skills need execution feedback and verification before becoming durable. Voyager stores executable skills after iterative prompting, environment feedback, execution errors, and self-verification. skill-miner candidate validation prompts, skill-personalizer workflow completion checks.
Public skill distribution requires provenance and safety review. Agent Skills survey highlights security concerns in community skills and argues for lifecycle governance. skill-generalizer redaction, platform compatibility, publication rubric.

Related Projects And Ecosystem

Project / Ecosystem Why It Matters
OpenAI Codex Agent Skills Official Codex support for skills, bundled scripts/references/assets, and plugin packaging.
Claude Code Skills Official personal/project/plugin skill scopes and description-driven loading.
Cursor Agent Skills Native Agent Skills support alongside rules and commands.
OpenCode Agent Skills Native skill tool and repo/home skill discovery.
Google / Gemini Agent Skills Google official skill repository and Gemini CLI / Antigravity-oriented skill installation.
Awesome Agent Skills / Antigravity Awesome Skills Community evidence that cross-agent SKILL.md packages are becoming a real distribution format.
CodeAlive Skills / GitHub Agent Skill projects Examples of skill packages paired with setup, references, scripts, and cross-agent install instructions.

Implications For This Project

  1. Keep lifecycle stages separate. Research on externalized memory and skill libraries supports generation from experience, but distribution and personalization have different safety and portability constraints.
  2. Audit routing, not just content. A skill can be well-written but useless if it undertriggers; trigger fit, overtrigger, and conflict checks deserve their own audit surface.
  3. Make mining configurable. Workflow discovery should use editable pattern definitions because user histories and agent log formats differ.
  4. Treat summaries as hints. Rollout summaries and compressed memories are valuable for recall, but raw transcripts or concrete examples should confirm candidates before creating a new skill.
  5. Use scripts for deterministic scanning. Repeated parsing, counting, and sanitization should live in scripts rather than being rewritten ad hoc by agents.
  6. Package for the target platform honestly. Cross-agent SKILL.md portability is real, but UI metadata, permissions, global paths, and plugin packaging differ by agent.

Papers And References

  • Memento-Skills: Let Agents Design Agents, arXiv:2603.18743.
  • Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward, arXiv:2602.12430.
  • Voyager: An Open-Ended Embodied Agent with Large Language Models, arXiv:2305.16291.
  • Reflexion: Language Agents with Verbal Reinforcement Learning, NeurIPS 2023 / arXiv:2303.11366.
  • ExpeL: LLM Agents Are Experiential Learners, arXiv:2308.10144.
  • Lost in the Middle: How Language Models Use Long Contexts, TACL 2024 / arXiv:2307.03172.
  • RAG-MCP: Mitigating Prompt Bloat in LLM Tool Selection via Retrieval-Augmented Generation, arXiv:2505.03275.
  • Model Context Protocol Tool Descriptions Are Smelly, arXiv:2602.14878.