9 principles that make agent skills actually work: selection, behavior change, decay
bibryam · x · 2026-09-11
A long-form guide distills nine principles for writing useful agent SKILL.md files, organized by lifecycle:
Act 1: Getting selected
- Metadata is the gate: at session start the model only sees the skill listing; the description is the primary selection signal, not a summary. State what the skill does, which requests trigger it, and which nearby work should not. Lose the metadata-only round and the body can't rescue you.
- Disclose progressively: context is a shared budget across skills.
Act 2: Reliably changing behavior — loaded skills must give executable, verifiable instructions.
Act 3: Contact with reality — skills that work today can quietly decay tomorrow and need ongoing maintenance.
Core thesis: a beautifully written Markdown essay is not a useful skill.
More from coding & agent
- Two days of heavy Cursor use: grok 4.6 xhigh handles routine tasks well — mazzaTalk · 2026-09-11
- AI agent edits professional videos via DaVinci Resolve's native MCP, all on one Mac mini — _AustinCalvert_ · 2026-09-11
- Astra Light shines at local Codex chores: file cleanup and installs, fast and capable — mark_k · 2026-09-11
- Analyst report: Agentic AI is cyber defense's next frontier as agents outpace governed production — BenBajarin · 2026-09-11
- Your coding agent forgets its rules — 7 research-backed fixes — victor_explore · 2026-09-11
- 10 free GitHub repos for AI agents and coding workflows worth bookmarking — Div_pradeep · 2026-09-11