victorbuto Slice Search Skill
5 steps: receive question → clarify → plan queries → approve → 3 parallel subagents search & analyze → merge.
Phase 0: Receive the Research Question
IMPORTANT: Do NOT ask any questions until the user has provided their research topic/question. If the user's message already contains a clear research question — proceed to Phase 1 (SPICE). If the message is vague or just triggers the skill without a specific question, respond with:
Какой вопрос будем исследовать?
STOP here and wait for the user's research question before proceeding.
Phase 1: Clarify the Research Question (SPICE)
Once you have the user's research question, ask 5 clarifying questions using the SPICE framework (adapted from academic research methodology: Setting, Perspective, Interest, Comparison, Evaluation).
The goal: expose blind spots and hidden assumptions in the user's request.
Ask these 5 questions and wait for answers before proceeding to Phase 2:
| # | SPICE element | Question to ask | What it reveals |
|---|---|---|---|
| 1 | Setting | Какая география, отрасль, сегмент рынка? Есть ли временные рамки (последний год, 5 лет, всё время)? | Boundaries — сужает область поиска |
| 2 | Perspective | Чей взгляд важен — бизнес, академия, потребители, инвесторы? Для какой аудитории результат? | Angle — определяет тип источников и язык |
| 3 | Interest | Что конкретно нужно найти — обзор темы, конкретные цифры/данные, кейсы, список игроков, тренды? | Depth — задаёт глубину и формат поиска |
| 4 | Comparison | Есть ли с чем сравнивать? Нужен бенчмарк (другая страна, период, конкурент)? | Contrast — добавляет сравнительное измерение |
| 5 | Existing knowledge | Что уже знаете по теме? Какие источники/статьи уже видели? Что НЕ нужно искать? | Gaps — исключает дубли, фокусирует на белых пятнах |
Format for asking:
Перед поиском уточню 5 вопросов по фреймворку SPICE:
1. **Рамки:** [вопрос адаптированный к теме]
2. **Перспектива:** [вопрос]
3. **Что ищем:** [вопрос]
4. **Сравнение:** [вопрос]
5. **Уже знаете:** [вопрос]
After user answers, incorporate their responses into Phase 2 analysis as constraints.
Phase 2: Plan Search Queries
Use SPICE answers to inform query generation.
2.1 Analyze the Query
- Core topic — main subject (from user request + SPICE answers)
- Goal — facts, sources, opinions, data, comparison (from SPICE #3 Interest)
- Query type — classify for slice distribution
- Constraints — from SPICE #1 Setting + #2 Perspective + #5 Existing knowledge
| Query Type | Primary Slices | Secondary Slices |
|---|---|---|
| Factual/Statistical | Source, Authority, Peristatic | Regional, Syntactic |
| Technical/Expert | Specific Term, Metonymical, Source | Syntactic, Authority |
| Current Events | Source, Regional, Authority | Peristatic |
| Historical/Academic | Syntactic (filetype), Authority, Source | Regional, Metonymical |
| Comparative | Regional, Source, Authority | Metonymical |
| Niche/Obscure | Specific Term, Syntactic, Peristatic | Regional, Source |
2.2 Generate 20+ Search Queries
Use 7 slice types (see references/slice-method.md):
| Slice | What It Does |
|---|---|
| Regional | Other languages/markets |
| Authority | Known institutions/experts |
| Source | Specific platforms/databases |
| Specific Term | Niche expert terminology |
| Syntactic | Alt spellings, scripts, filetypes |
| Metonymical | Adjacent/related concepts |
| Peristatic | Quotes, contextual fragments |
Key: Rare terms > generic terms.
Distribution by type:
| Query Type | Reg | Auth | Source | Spec | Synt | Meton | Perist |
|---|---|---|---|---|---|---|---|
| Factual | 2 | 3 | 8 | 2 | 1 | 1 | 3 |
| Technical | 1 | 2 | 4 | 5 | 3 | 4 | 1 |
| News | 3 | 3 | 9 | 1 | 1 | 1 | 2 |
| Academic | 2 | 4 | 4 | 2 | 4 | 2 | 2 |
| Comparative | 5 | 3 | 6 | 2 | 1 | 2 | 1 |
| Niche | 3 | 1 | 1 | 5 | 4 | 3 | 3 |
2.3 Present to User
STOP. Show the query list and wait for approval:
## Search Plan: [Topic]
**Type:** [Classification] | **Queries:** [N]
| # | Slice | Query | Why |
|---|-------|-------|-----|
| 1 | Source | `site:statista.com ...` | Stats |
| ... | ... | ... | ... |
Approve, edit, or add?
Phase 3: Search & Analyze (3 parallel subagents)
After user approves, split queries into 3 groups and launch 3 subagents in parallel. Each subagent searches AND analyzes within its own context window.
Subagent prompt template
Task(subagent_type="general-purpose", prompt="
You are a research agent. Your task:
1. SEARCH: Execute these queries using tavily_search.
For each query: search_depth='advanced', include_raw_content=true, max_results=5
Queries:
1. [query]
2. [query]
...
2. ANALYZE: Read through all results (snippets + raw_content).
Extract: facts, statistics, expert quotes, case studies, contradictions.
Drop results with score < 0.3. Deduplicate by URL.
3. RETURN exactly this format:
## Round [N] Results
**Queries executed:** [N]
**Results found:** [N] unique after dedup
### Key Findings
[Every fact, statistic, or claim with inline source link:
'fact or data — [Source Name](URL)']
### Source List
| # | Title | URL | Score | What it contributed |
|---|-------|-----|-------|---------------------|
### Gaps
[What queries returned nothing useful, what's missing]
")
Launch all 3 in parallel in a single message:
- Subagent 1: queries 1-8 (priority slices)
- Subagent 2: queries 9-15 (secondary slices)
- Subagent 3: queries 16-22+ (adaptive, gap-filling)
What main context receives
3 structured summaries (~3-5KB each) with pre-analyzed findings and inline sources. NO raw page content in main context.
Phase 4: Merge & Synthesize
Combine findings from all 3 subagents into final output.
4.1 Merge
- Combine Key Findings from all rounds
- Deduplicate facts (same stat from different sources → keep best source)
- Note contradictions between rounds
4.2 Output
Source attribution rule: Inline hyperlink right after the fact it supports: fact — [Source Name](URL). Do NOT collect sources into a separate table at the bottom.
## Research: [Topic]
**Queries:** [N] | **Sources found:** [N] | **Rounds:** 3
### Key Findings
[Synthesized answer organized by themes/subtopics.
Each fact, statistic, or claim has an inline source link:
"fact or data — [Source Name](URL)"]
### Search Effectiveness
| Metric | Details |
|--------|---------|
| **Best slices** | [slice type] — [why it worked, example result] |
| **Failed slices** | [slice type] — [why it failed or returned noise] |
| **Gaps** | [what's missing] |
| **Follow-up queries** | `query 1`, `query 2` |
Quick Reference
| Slice | Example |
|---|---|
| Regional | AI news → actualités IA France |
| Authority | climate → IPCC report findings |
| Source | AI → site:arxiv.org transformer |
| Specific Term | coffee → anaerobic fermentation |
| Syntactic | report → filetype:pdf quarterly |
| Metonymical | blockchain → Byzantine fault tolerance |
| Peristatic | article → "absorbs 30 percent" |