prompt-refinement

"[guide] Extract behavior principles from external

Prompt Refinement

Distill actionable behavior principles from external sources into personal agent prompts.

Boundary

  • Does:
    • Analyze external projects for agent behavior patterns.
    • Evaluate which patterns apply to the user's agent context.
    • Distill principles into concise, rhetoric-free rules.
    • Write principles into target prompt files.
  • Does not:
    • Write Fungus module SKILL.md — skill-spec handles that.
    • Teach general prompt engineering theory.
    • Modify Kiro system prompt or built-in behavior.

Interface

  • Triggers: user shares an external project to evaluate, asks to improve agent prompt behavior, mentions agent-prompt.md or work prompt refinement.
  • Related: skill-spec for Fungus skill authoring, fruit for skill emergence from nutrients.

Behavior

Evaluate External Source

  1. Situation — user shares a repo, framework, or methodology and asks what is useful.
  2. Guidance —
    • Read the source material thoroughly.
    • Identify behavior-shaping patterns: role definitions, decision heuristics, quality gates, failure responses.
    • Classify each pattern:
      • Applicable — addresses a gap in current prompts.
      • Already covered — current prompts handle this.
      • Not relevant — wrong domain or abstraction level.
    • Present the classification with reasoning. Let the user decide what to adopt.
  3. Handoff — if the source contains Fungus module ideas, hand off to skill-spec.

Refine Prompt

  1. Situation — user confirms which principles to adopt, or asks to improve a specific prompt file.
  2. Guidance —
    • Read the target prompt file first.
    • Strip rhetoric, jargon, and pressure language from the source principles. Keep only the actionable core — the agent is smart enough to follow reasoning without being threatened.
    • Write principles as short imperative phrases. Each principle should be one line, self-contained.
    • Group under a clear heading (e.g. 行为原则).
    • Verify: would a fresh agent session understand and follow each principle without extra context?
  3. Handoff — if the user wants to test prompt changes across sessions, suggest manual verification in a new chat.

Quality Check

  1. Situation — after writing principles, self-review.
  2. Guidance —
    • Each principle must be actionable (starts with a verb).
    • No overlap — two principles should not say the same thing in different words.
    • No vague aspirations — "be proactive" is too vague; "发现隐患主动提出，不等用户追问" is actionable.
    • Total principles per prompt should stay under 7. Beyond that, the agent stops internalizing them.
  3. Handoff — none. This is the final step.