S-TIER repo
Microsoft SkillOpt — agent skills that rewrite themselves against benchmarks
100k ft — What is this?
Microsoft SkillOpt — agent skills that rewrite themselves against benchmarks
50k ft — Why care?
Microsoft open-sourced SkillOpt: a base model runs a task, an optimizer rewrites the skill instructions from results, successful paths are separated from failures, and edits that don't beat the benchmark are auto-rejected. Claimed to beat hand-written prompts and TextGrad and to transfer across models.
10k ft — Key takeaways
• Gate every rewrite on beating the benchmark.
• Model-agnostic, so optimized skills port across providers.
• Treat skill instructions like trainable parameters with an eval set.
• link: https://github.com/microsoft/SkillOpt
5k ft — Tools & techniques
SkillOpt
AI agents, automatic prompt optimization, benchmark-gated edits
1k ft — Deep-dive / try it
🚨 MICROSOFT OPEN-SOURCED SELF-EVOLVING AGENT SKILLS
SkillOpt lets skills improve themselves. A base model runs the task, an optimizer rewrites the instructions based on results.
Ships with:
• Isolates successful paths from failures
• Auto-rejects edits that don’t beat the benchmark
• Beats hand-crafted prompts and TextGrad
• Zero lock-in, transfers to any model
100% free. Open source.
👉 https://github.com/microsoft/SkillOpt
Who’s evolving their agent skills today? Drop a 🔥
Transcript excerpt
What's up guys, so Microsoft just open-sourced self-evolving agent skills Okay skills that improve themselves the same way you train an AI model So no more guessing whether your prompt tweaks helped again very fun open source hundred percent free Someone can use this It optimizes the outputs and evaluates them and rewrites the instructions themselves, you know It's ages. There's agent skill builder, dude. Come on now
Appears in
- Guide: Skills That Earn Their Lines — benchmark-gated self-editing plus subtraction
- Deeper dive: Skills That Earn Their Lines — benchmark-gated self-editing plus subtraction
- Guide: Strategy Autoresearch — the judge is the product, not the agent
- Deeper dive: Strategy Autoresearch — the judge is the product, not the agent
- Guide: Benchmark Your Brain — the skill-optimizer loop, pointed at the human
- Deeper dive: Benchmark Your Brain — the skill-optimizer loop, pointed at the human
More from @marc.kaz
- B-TIER ASCILINE — 30fps video streamed as colored text over WebSockets
- S-TIER PPF Contact Solver — ZOZO's open-source GPU cloth and soft-body physics
- S-TIER Blender — • 500+ body, hand & face landmarks per frame
- A-TIER Carbonyl — full Chromium browser rendered inside your terminal
- A-TIER PLFM_RADAR — open-source radar with PCBs, FPGA code and Python GUI