cobusgreyling

Skill Boundary — AI skill for Claude Code

AI community

AI Agents & The Skill Boundary Problem — exploration-first execution and skill boundary awareness based on the OSExpert paper.

How to install Skill Boundary

This entry records only its repository, not the path inside it, so there is no exact command to give. Open cobusgreyling/skill-boundary and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

What Skill Boundary does

AI Agents & The Skill Boundary Problem — exploration-first execution and skill boundary awareness based on the OSExpert paper.

Alternatives in AI

Oh My Claudecode — 32 specialized agents and 7 execution modes for Claude Code, with smart model routing and automatic paralleliz 10.9k ★
Open Multi Agent — TypeScript multi-agent orchestration engine — one runTeam() call from goal to result 5.8k ★
Welcome — AI Research Skills — You now have access to 86 production-ready skills covering the entire AI research lifecycle: literature survey 5.4k ★

README

AI Agents & The Skill Boundary Problem

![Header](header.jpg)

AI Agents that do not know what they cannot do waste most of their compute trying. Based on the [OSExpert paper](https://arxiv.org/abs/2603.07978v1) by Liu et al. — agents that explore an environment first and map their skill boundaries achieve ~20% performance gains and ~80% efficiency improvements.

What is in this repo

blog.md — Full blog post on skill boundary awareness, exploration-first execution, and universal agency
explore-and-map.py — AI Agent explores an unknown developer platform API, discovers capabilities, and maps skill boundaries using NVIDIA Nemotron 3 Super
skill_boundary_heatmap.py — Generates the skill boundary heatmap visualisation
exploration_report.md — Full agent exploration output
discovered_api_map.json — Structured API discovery data

Skill Boundary Heatmap

![Skill Boundary Heatmap](skill_boundary_heatmap.png)

Key findings from the paper

Agents that explore before executing achieve ~20% higher success rates
The skill boundary check (stopping early on known failures) provided most of the efficiency gain
General-purpose agents spend 5-50x longer than human experts — most of it wasted on impossible tasks
Recording failures is as valuable as recording successes