answer-tree-auto

name	answer-tree-auto
description	Use when the user asks follow-up questions about a previous long LLM answer, references a section/point/node in Answer Tree, or asks for a new long answer that should be saved. Automatically decide whether to call answer_tree tools to create, prompt, attach, or skip nodes.

Answer Tree Auto

Use this skill to keep Answer Tree updated during normal chat. The user should be able to ask naturally; do the tool calls yourself when the message should become part of the tree.

Goal

Maintain multiple trees inside one OpenCode session:

session
  root topic A
    -> follow-up answer saved as child node
  root topic B
    -> follow-up answer saved as child node

There is only one active node/context at a time. The user can switch focus between topics in the sidebar; follow-up questions attach to the active node only when the user clearly refers to that context.

Parent routing rule:

• For each new user question, first decide which existing node should be the parent context.
• The active node is the default parent.
• Route to another node when the user clearly names another node id, title, section, topic, or source question shown in the sidebar.
• If another node is only weakly related, keep the active node.

Depth rule:

• The tree should have at most 3 levels: 1 -> 1.1 -> 1.1.1.
• Never create or attach a child node under a third-level node.
• For fourth-level or deeper follow-ups, answer normally inside the current context.

Prefer quiet automation. Do not ask the user to manually call answer_tree_* tools unless a required tool is unavailable or the target node is ambiguous.

Output discipline:

• Never print internal planning, reasoning, or workflow narration.
• Never start a visible message with Thought:.
• Never explain that you are about to call a tool.
• Never output the same answer twice.
• The visible response should contain only:
  1. the final answer once
  2. one short Answer Tree status line when a tool was used or intentionally skipped

Decision Rules

Track the user message when at least one is true:

• The user references a previous long answer, section, bullet, paragraph, point, or node.
• The user says things like "针对这一部分", "展开这个点", "第二点", "用户体验部分", "继续追问", "上面那段", "这个回答里".
• The user asks a follow-up that is likely to produce a new reusable answer node.
• The user asks for a new standalone long answer that they may later ask multiple follow-ups about.

Do not track when:

• The message is a greeting, typo fix, UI operation, install/run/debug question, or short clarification.
• The user explicitly says not to save or not to update Answer Tree.
• The answer will be short and unlikely to become a future branch.
• You cannot infer which existing node the question belongs to and there is no active node.
• The user asks a very narrow detail question, such as a single definition, one equation, one number, one concrete output, one example, or "why is this value X".

New-topic rule:

• Treat a fresh topic as a new root node even when an active node exists.
• Fresh-topic examples: "what is CNN", "解释 CNN 是什么", "介绍一下 RAG", "对比 CNN 和 Transformer", "新话题".
• Only attach under the active node when the user clearly says "继续", "这个", "这里", "上面", "刚才", "第 2 点", "当前节点", "this section", "above", "previous", or names a visible node/section.

When unsure whether the user changed topic, prefer a new root node. Use the active node only when the message has a clear follow-up reference.

Workflow A: Follow-Up On Current Active Node

Use when the user asks a clear follow-up about the currently selected/active answer.

1. Call answer_tree_prompt_auto with the user's question.
2. Read the returned context path and draft the answer silently. Do not print the draft yet.
3. If the drafted answer is substantial enough to be useful as a future node, call answer_tree_attach_last with:
   • content: the exact assistant answer content you will later show the user
   • title: a short, specific title, preferably 6-14 words
4. Final response should include the answer exactly once. Add one short Answer Tree status line.

Skip step 3 when the answer is short, procedural, or not worth future follow-up.

Hard rule:

• Never call answer_tree_attach_last before answer_tree_prompt_auto, answer_tree_prompt_current, or answer_tree_prompt in the same follow-up flow.
• Prefer answer_tree_prompt_auto for natural follow-up questions, because it chooses the parent node before recording the question.
• Never put the user's question, feedback, or request text into answer_tree_attach_last.content.
• Phrases like "看不懂", "解释一下", "展开第三步", "这里再详细讲" are user questions. First record them with answer_tree_prompt_auto, then answer, then attach the assistant answer if substantial.
• For very specific detail questions, call answer_tree_prompt_auto when context is needed, answer normally, and do not call answer_tree_attach_last.
• If the active node is already third-level, do not call answer_tree_attach_last.

Workflow B: New Standalone Long Answer

Use when the user asks for a fresh long answer that is not a follow-up to an existing node.

Use this workflow even if an active node exists, as long as the user asks about a different topic.

1. Draft the answer silently. Do not print the draft yet.
2. If the drafted answer is long or structured enough to become a root node, call answer_tree_create with:
   • content: the exact assistant answer content you will later show the user
   • title: a concise topic title
3. Final response should include the answer exactly once. Add one short Answer Tree status line.

Good root candidates:

• multi-section product/technical analysis
• plans, architecture, comparisons, research summaries
• long explanations with several future follow-up points

Bad root candidates:

• single command
• one-paragraph answer
• transient debugging advice
• status updates

Workflow C: User Names A Specific Node

Use when the user mentions a node id or asks to base a question on a visible node.

1. If the node id is explicit, call answer_tree_prompt with that nodeId and the question.
2. If the user asks to switch to a node by title or topic, call answer_tree_use_match with that title/topic before answering.
3. If the user only describes the node title/section and there is ambiguity, call answer_tree_use_match first; if the result is still wrong or unclear, call answer_tree_list or ask one concise clarification.
4. Draft the answer silently using the returned path context.
5. If substantial, call answer_tree_attach_last with the exact answer content and title.
6. Final response should include the answer exactly once. Add one short Answer Tree status line.

Again, answer_tree_attach_last is only for the assistant's answer after a question has already been recorded. It should not be the first Answer Tree tool in a follow-up.

What Counts As Substantial

Attach the assistant answer as a child node when it has one of these shapes:

• at least 2 paragraphs
• at least 3 bullets or numbered steps
• introduces a new plan, design, explanation, comparison, or decision
• likely to be asked about again

Do not attach for answers that are only "yes/no", one command, one sentence, a small correction, a single example, a single formula, a concrete numeric explanation, or a local clarification.

Tool Preference

Prefer current-node tools:

answer_tree_prompt_auto
answer_tree_prompt_current
answer_tree_attach_last

Use explicit-node tools only when needed:

answer_tree_prompt
answer_tree_show
answer_tree_use
answer_tree_use_match

Use creation/export/list tools when the user's intent matches:

answer_tree_create
answer_tree_list
answer_tree_export

Use management tools only when the user explicitly asks:

answer_tree_show_full
answer_tree_rename
answer_tree_delete

User-Facing Behavior

Keep the response clean for live demos. Do not expose reasoning or workflow mechanics. Show tool status only as a short final line.

Good final shape:

<formal answer, shown once>

Answer Tree: saved as a child node via answer_tree_prompt_auto + answer_tree_attach_last.

When a new root is saved:

Answer Tree: saved as a root node via answer_tree_create.

When the answer is intentionally not saved:

Answer Tree: not saved; this was a narrow detail answer.

Keep the status line short. The main answer still matters most.

Failure Handling

If a needed Answer Tree tool is unavailable:

• Answer normally.
• Say briefly that Answer Tree auto-save did not run because the tool is unavailable.

If the tool call fails:

• Answer normally if possible.
• Mention the failure briefly and avoid retry loops.

If the active node is wrong:

• Use the user's explicit node/section if clear.
• Otherwise ask the user to select the right Answer Tree node first.