spread1000-context-collector

name	spread1000-context-collector
description	Collects missing context from the user via one-question-at-a-time dialogue, generates a 6-element meta-prompt, and passes it to downstream skills. Applies shikigami's interactive purpose-discovery pattern specialized for SPReAD. Use when user request is ambiguous (依頼が曖昧), research field is unknown (研究分野が不明), or required information is insufficient (必要情報が不足).

Context Collector

Assess context sufficiency from the user's prompt, collect missing information via one-question-at-a-time dialogue, and generate a 6-element meta-prompt to maximize downstream phase accuracy.

Use This Skill When

• The user's request is ambiguous and it is unclear which skill to route to
• Any of research field, theme, or purpose is unclear
• Required Inputs for downstream skills (e.g., research-planner) cannot be satisfied

Context Sufficiency Check

Determine whether the following 6 elements can be extracted from the user's initial prompt. If 3 or more elements are unknown, activate this skill.

Element	Meaning in SPReAD	Criteria
PURPOSE	Research goal / what to achieve with AI	Contains action verbs like 「〜を解明」「〜を予測」
TARGET	Research subject / field	A specific academic field or subject is explicitly stated
SCOPE	Scope / scale of research	Data scale, target period, or geographic scope is mentioned
TIMELINE	Research period	Awareness of the ~180-day research period
CONSTRAINTS	Constraints	Budget, equipment, personnel, or ethical constraints are mentioned
DELIVERABLES	Expected outputs	Papers, software, models, etc. are mentioned

Workflow

Step 1: Context Sufficiency Assessment

Analyze the user's initial prompt and internally assess the status of the 6 elements.

判定結果の例:
  PURPOSE:      ✅ 「材料探索を加速したい」
  TARGET:       ✅ 「無機固体材料」
  SCOPE:        ❌ 不明
  TIMELINE:     ❌ 不明
  CONSTRAINTS:  ❌ 不明
  DELIVERABLES: ❌ 不明
  → 不足4要素 ≥ 3 → context-collector を起動

Step 2: One-Question-at-a-Time Information Collection

Rules:

• Ask only 1 question per message (simultaneous multiple questions are prohibited)
• Complete collection in a minimum of 3 and a maximum of 7 questions
• Attach a category label to each question
• If the user answers "わからない" (don't know), propose an estimated value and confirm

Question Categories and Order:

Order	Category	Example Question (SPReAD-specific)
1	WHY	この研究で AI を使って何を達成したいですか？
2	TARGET	研究対象の分野・物質・現象は何ですか？
3	DATA	どのようなデータを扱いますか？（種類・規模・形式）
4	TIMELINE	SPReAD の研究期間は約180日間（6ヶ月）ですが、この期間で取り組む範囲として想定していることは何ですか？
5	CONSTRAINT	予算・設備・人員の制約はありますか？（SPReAD の予算上限は直接経費500万円です）
6	SUCCESS	どのような成果が得られれば成功ですか？
7	EXISTING	現在の研究手法で課題に感じていることは何ですか？

Question Format:

## ❓ 質問 N/M
**カテゴリ**: WHY
この研究で AI を使って何を達成したいですか？
（例: 新規材料の候補を網羅的にスクリーニングしたい、気象シミュレーションの精度を向上させたい）

⚠️ In CONSTRAINT category questions, the example text must accurately state SPReAD の予算上限は直接経費500万円（間接経費を含め最大650万円）. Never state incorrect amounts such as 「1,000万円」.

Step 3: Meta-Prompt Generation

Structure the collected information into a 6-element meta-prompt.

• Reuse assets/meta-prompt-template.md when producing the meta-prompt

Meta-prompt Display Format:

## 📋 構造化メタプロンプト

| 要素 | 内容 |
|------|------|
| PURPOSE | [収集した目的] |
| TARGET | [研究対象・分野] |
| SCOPE | [範囲・規模] |
| TIMELINE | [期間（約180日間）] |
| CONSTRAINTS | [制約条件] |
| DELIVERABLES | [期待成果物] |

この内容で研究プラン策定を進めてよろしいですか？
修正があればお知らせください。

Step 4: User Approval

• Display the meta-prompt and always wait for user approval / ユーザー承認
• If modifications are requested, update the relevant elements and re-display
• 承認を得た後 (after approval), save the meta-prompt to output/{project-name}/meta-prompt.md

Step 5: Handoff to Downstream Skills

Based on the approved meta-prompt, hand off processing to the optimal skill following the routing rules in AGENTS.md. The meta-prompt content becomes the input context for the downstream skill as-is.

Deliverables

• output/{project-name}/meta-prompt.md: Approved 6-element meta-prompt

Quality Gates

• All 6 elements are filled (「不明」/「未定」 have been supplemented with estimated values)
• Meta-prompt is finalized only after obtaining user approval
• Multiple questions are never asked simultaneously in a single message
• A minimum of 3 dialogue turns have been conducted

Gotchas

• Even if the user says 「全部お任せ」 (leave everything to you), explicitly confirm at least PURPOSE and TARGET. Proceeding on estimates alone causes major rework downstream
• Question order is not fixed. Skip elements already known from the user's initial prompt
• When receiving a 「わからない」 response, propose typical values for the research field and confirm
• Never ask about the application deadline in TIMELINE questions. The 公募期間 is fixed at 「令和8年4月17日～5月18日正午」. The research period is also fixed at ~180 days. TIMELINE confirms the scope and milestones for the 180-day period

Validation Loop

1. Assess the sufficiency status of the 6 elements
2. Check:
   • All elements are filled with concrete values
   • User approval has been obtained
   • Required Inputs for downstream skills can be satisfied
3. If any check fails:
   • Ask additional questions for insufficient elements
   • Propose estimated values and confirm
4. Finalize the meta-prompt only after all gates pass