Dex Horthy 谈多 Skill 同会话是否污染上下文及取舍方法

dex· @dexhorthy · X·2026-10-26 01:27· 59 分钟前AI 评分52

AI 导读

HumanLayer 的 Dex Horthy 回应提问，认为同会话串联约 5 个 Skill 没问题，但 Skill 需要叠加生效或数量极多（如 200 个）时，应改用子智能体、新会话或把多个 Skill 内容预合并到一条用户消息。他建议按内容类型取舍：流程步骤宜拆开，纯信息可合并，重要行为规则需重复注入或用 hooks 等确定性手段保证遵循，并提醒不要把权重里已有的知识写进 Skill。

正文 · 原文

question from the TL:

quick question, do you think consuming more skills in the same session unnecessarily dirties the context?

Say you have a 5 step work, each step with a skill. You can either run all on single session, divide into subagents with relevant skills, or compact or new chat per step.

my intuition says 5 steps in a process is fine. Each successive skill will take a share of attention from previous skills,

so if you have a VERY long complex process where you need skills to STACK (ie instructions from skill 2 must still be honored while skill 6 is executing)

Or if you have many (eg 200) separate skills that all need to be active during a session ( think: agent memory snippets implemented as skills, probably an antipattern in today’s harnesses)

Then you want fresh context, subagents, or to synthetically precombine many skill contents into a single user message (although then you lose the attention recency bias for instructions in step 6 of the process)

Also worth noting that good skills can be a mix of:

- information not in the weights
- steps to execute a process (not in the weights)
-general steering or behavior

They usually contain all three, but what’s in your skills nudges the meta here.

More instructions / process steps? You wanna string those out

Mostly information? Probably safe to combine into one user message.

Behavior? Depends. Adherence can decay over a long session. Fine to split or combine, but if it’s important, you want some way of repeating important rules, either by including rules in all the steps of a process, or by something more deterministic like hooks

And OF COURSE I’m begging you PLEASE stop putting stuff that’s in the weights in your skills. You are wasting time and locking in a specific, potentially suboptimal behavior.

来源：dex · x.com