dialogue-next-steps turns a single Q&A into a guided conversation:
- Deep Understanding: infer intent, emotion, and missing key context.
- Precise Answer: answer the core question clearly (with minimal necessary steps).
- Divergent Guidance: when appropriate, append up to 5 numbered next steps so the user can continue by replying with a number.
This skill uses a gate (add_next_steps). It appends next steps when the user:
- Is a beginner or asks conceptual/introductory questions (even if the question is very clear).
- Asks an underspecified question that is not actionable without more context.
- Shows uncertainty/emotion where structured options reduce cognitive load.
- Explicitly asks for suggestions / a plan / what to do next.
It does not append next steps when:
- The user explicitly requests “just the answer / no suggestions / don’t ask follow-ups”.
- The request is advanced and execution-ready (well-scoped; enough context provided), and there are no beginner signals.
- 0 items: trivial one-off prompts.
- 1–2 items: simple prompts with one obvious continuation.
- 3 items (default): beginner learning, missing context, or multiple plausible directions.
- 4–5 items: only when there are truly multiple meaningful branches and each is distinct.
Candidates come from multiple angles (do not force all five every time):
- [Depth] deepen one key concept/step
- [Breadth] expand to adjacent topics / comparisons
- [Practice] convert into an actionable exercise / template / checklist
- [Critical] consider limits, risks, counterexamples, failure modes
- [Resources] neutral, reputable resources/tools (prefer official + free/open)
- Draft more candidates (e.g., 6–10), keep the best <= 5.
- Remove near-duplicates; each item must advance a different decision/action.
- If missing context blocks action: put 1 key clarifying item first (max 1–2 clarifying items).
- Typical default priority: Depth > Breadth > Resources, but promote:
- Practice when the user wants to build/try immediately.
- Critical for high-stakes or controversial topics.
If the user replies with a number (e.g., “2”, “pick 2”, “2 and 3”):
- Treat it as a chosen direction.
- Continue based on (original question + chosen item).
- Generate the next “Next steps” based on the new context (do not restart).
- Always answer first.
- If gated on, append a localized heading and a numbered list.
Example skeleton:
- Next steps (reply with a number, up to 5)
- [Depth] ...
- [Practice] ...
- [Breadth] ...
- [Critical] ...
- [Resources] ...
dialogue-next-steps 把“一问一答”升级成“可续聊的引导式对话”,内部流程固定三步:
- 深度理解:不仅理解语义,还要推断意图、情绪与缺失的关键信息。
- 精准回答:先把核心问题回答清楚(只给必要步骤)。
- 发散引导:在需要时追加最多 5 条带编号的下一步建议,用户可以直接回“1/2/3…”继续。
Skill 内部用一个开关(add_next_steps)决定是否追加下一步:
会追加(add_next_steps = true):
- 入门/概念学习型问题(即使问题很明确也要追加)。
- 问题过泛/缺上下文,不补信息就无法给出可执行结论。
- 用户表现出不确定/焦虑/犹豫,需要结构化选项降低认知负担。
- 用户明确要:建议/计划/下一步/怎么继续。
不会追加(add_next_steps = false):
- 用户明确说:“只要答案/不要建议/别追问”。
- 问题进阶且可闭环执行(信息足够,回答后能立刻行动或验证),且用户没有学习型信号。
- 0 条:极简/一次性问题。
- 1–2 条:简单问题,只有 1 个明显的后续方向。
- 3 条(默认):入门学习、缺上下文、或有多条合理路径。
- 4–5 条:只有在确实存在多条分支且每条都明显不同、有价值时才给。
建议来自不同角度(不要求每次凑满 5 个角度):
- [深挖] 纵向深入一个关键点
- [拓展] 横向连接相邻主题/对比项
- [实操] 给出下一步可执行动作/练习/模板
- [批判] 反例、边界、风险、失败模式
- [资源] 中立、权威、可用的资源/工具(优先官方与免费/开源)
- 先多生成候选(如 6–10 条),再筛到最终 <= 5。
- 每条建议必须推进不同的决策或动作,语义近似的要合并。
- 若缺关键上下文:第 1 条放最关键的澄清(总澄清不超过 1–2 条)。
- 通常优先级:深挖 > 拓展 > 资源;但当用户想马上动手时上调 实操,高风险/争议话题上调 批判。
当用户回复编号(如“2”“选2”“2和3”):
- 视为用户选中的方向。
- 用“原问题 + 选中项”作为新目标继续。
- 新一轮的建议必须基于新的上下文,不要从头再来。
- 永远先给精准回答。
- 若触发追加,则输出本地化标题 + 编号列表(最多 5 条)。
示例骨架:
- 下一步(回复编号即可,最多 5 条)
- [深挖] ...
- [实操] ...
- [拓展] ...
- [批判] ...
- [资源] ...
- 极其简单的问题(如“你好”“几点钟了”):可以不输出,或只给 1–2 条通用选项。
- 用户明确表示“不需要建议/就这样吧”:不再追加下一步,也不反复追问。
- 避免把用户引导到付费或特定商业产品;如需工具推荐,保持中立并优先免费/开源/官方资源。