AgentSkillsCN

cognitive-load-budget

通过分块处理与“认知负荷预算”的合理调控，有效降低指令遗漏的风险。当提示信息过长（包含大量约束条件、冗长的文本记录或多项任务组合）时，模型很可能因信息过载而忽略后续指令，此时可借助此方法加以优化。

SKILL.md
--- frontmatter
name: cognitive-load-budget
description: Reduce instruction drop by controlling prompt information density using chunking and a “cognitive load budget”. Use when prompts are long (many constraints, long transcripts, multi-part tasks) and the model may ignore late instructions.

Cognitive Load Budget

Rules

  1. •Split prompts into at most three blocks, each ≤ ~10 bullets:
    • •Block A (Role + objective): who the model is, what success looks like.
    • •Block B (Constraints + output schema): strict formatting and must/shall rules.
    • •Block C (Inputs): transcripts, policy text, domain description.
  2. •If inputs are large, require a 2-step run:
    • •Step 1: summarize inputs into a compact, structured representation.
    • •Step 2: produce the final artifact from the summary.
  3. •Prefer schemas over long prose:
    • •“Fill these fields” reduces drift more than “be thorough”.

Prompt snippet (copy/paste)

text
Block A — Role & objective (≤10 bullets)
Block B — Constraints & output schema (≤10 bullets)
Block C — Inputs (paste text)

If Block C is long: first summarize C into a table/bullets, then produce the requested output.