Use agent skills, but don't overload them: soft rules can't stop prompt injection
bendee983 · x · 2026-08-31
The author argues the real question isn't whether to use AI agent skills, but what belongs in them. Skills are useful for organizing instructions and avoiding context-window bloat in the underlying LLM. However, he sees more people putting rules and guardrails into skills that would be better enforced by deterministic code and hard measures like ACLs and permissions. Natural language instructions are at best soft rules — anything written in a prompt or skill can eventually be overwritten by prompt injection or confusing context. Plan for that: use skills properly but don't overuse them, or it becomes a nightmare.
More from coding & agent
- Hermes Agent with local Qwen3.8-27B performs perfectly on coding tasks — Teknium · 2026-08-31
- Agent security reviews delaying launches by a quarter — ReasonabloeBottle468 · 2026-08-31
- AI 'research assistant' vs 'research agent': beyond paper summarization — Senior_Weakness255 · 2026-08-31
- A2A Agent discovering and calling public MCP servers: results — wagerx · 2026-08-31
- StarHarness: evolving fixed-weight agent harnesses for enterprise tool use — ServiceNow-AI · 2026-08-31
- Dev: OpenClaw Replaces T3Code as Primary Agent Environment — steipete · 2026-08-31