Blog
Notes on agent-native workflow, prompt-composed systems, and the files your agents actually read.
2026-06-29
Instruction Bleed: when your SKILL.md files quietly interfere
A new arXiv paper names a real failure mode: edit one prompt module and an unrelated one's behavior shifts. What it measured — and what it doesn't claim.
2026-06-29
What Instruction Bleed means if you write SKILL.md / AGENTS.md
The practitioner's read on instruction bleed: where cross-module interference is most likely to bite, and the cheap check the paper hands you.