ARC AGI 3 — Skill Issue
ARC AGI 3 — Skill Issue
ARC AGI 3 demonstrates how new skills can enhance AI model performance, providing greater accuracy and consistency.
ARC AGI 3 — Skill Issue
Skills have proven to be much better than system prompts. They can embed tool use, addressing the issue of repeatability and consistency of responses. This is an incredibly useful feature when a strict output format is needed, say, for filling out tables.
The team, using a custom skill for Claude Opus 5, achieved a 100% score on ARC AGI 3. The skill literally forces the model to think twice before making predictions on the benchmark. As a result, the model jumped from 30% to a clean 100%, which is already better than previous results, while using about half the number of steps.
There was no cheating in the form of dataset rules within the skill. The tool call simply has a hard logical block embedded. If the model does not reason before answering and does not try to predict the next step by filling in the corresponding cells, the prediction is sent back for revision. Otherwise, it follows basic instructions.
On one hand, RL seems to have killed classic prompt engineering, and now skill engineering has firmly taken its place.
This also shows how ineffective the ARC AGI benchmark has been on our trajectory to AGI (no offense, François).
Why it matters
AnalysisThis is important as it illustrates how new approaches to AI training can significantly enhance their performance. The use of skills instead of traditional prompts opens up new possibilities for the development of artificial intelligence technologies.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Codex Plugins for Beginners: The Complete Guide to the New Skill Ecosystem
A comprehensive guide to the Codex plugin ecosystem: understanding Skills vs. Plugins, connecting Marketplaces, GUI and CLI management, MCP server authorization, and packaging custom plugins.
How to Build Custom Skills for Claude Code: Architecture and Best Practices
A comprehensive engineering guide to creating custom Claude Code skills: folder structure and SKILL.md, progressive disclosure, skills vs. hooks trade-offs, and evaluation-driven workflow development.
Agent Skills & Custom Workflows
An architectural pattern for dynamically loading specialized procedural instructions, scripts, and templates (SKILL.md) into an agent's context window on demand (On-Demand Loading).
Thinking Machines Releases Inkling 2 Small: a 594B MoE Model That Can Fine-Tune Other Models
Thinking Machines released Inkling 2 Small, a 594B-parameter MoE model that leads among open models in several benchmarks and demonstrated autonomous fine-tuning of Qwen3.5-4B.