Overview
This skill uses the Evaluator-optimizer pattern for systematic multi-round web research, using a structured Ask process to clarify requirements before searching. Core features:
- Three-layer Ask framework — research goal classification → mode-specific questions → YAML requirements document
- Smart skip rules — automatically skip questions already answered based on user input
- Escape hatch quick mode — provide controllable degradation options when the user is impatient
- Research mode classification — focused/scan/comprehensive three modes correspond to different strategies
- Evaluator-optimizer loop — quality-driven iterative search
工作流程:
1. 需求澄清 — 三层 Ask 框架
├─ 1.1 项目重述(Re-ground)
├─ 1.2 研究目标分类
├─ 1.3 模式专用提问(聚焦/扫描/全面)
├─ 1.4 逃生舱口(Escape hatch)
└─ 1.5 输出研究需求文档(YAML)
2. 评估任务与查询构建
├─ 2.1 确定搜索模式(easy/medium/hard)
├─ 2.2 选择初始研究方向
├─ 2.3 创建搜索任务(todos)
└─ 2.4 构建查询集群(运用高级搜索技巧)
3. 分发子代理搜索(并行)
├─ 3.1 创建搜索团队(TeamCreate)
├─ 3.2 启动并行搜索(2~4 个 Agent)
└─ 3.3 子代理写入独立文件(turn-{n}/search-agent-{n}-task.md)
4. 聚合结果到结构化日志
├─ 4.1 初始化会话目录(/tmp/search-web/{date}/{task}/)
├─ 4.2 创建协调日志(master-log.md)
└─ 4.3 父代理记录事件时间线
5. 评估器 - 优化器循环
├─ 5.1 启动评估器(读取 turn-{n}/search-agent-*.md)
├─ 5.2 评估器输出报告(turn-{n}/search-evaluator-turn-{n}.md)
├─ 5.3 决策点(全 5/5 则 COMPLETE,否则 CONTINUE)
└─ 5.4 启动优化器(生成 turn-{n}/query-optimizer-turn-{n}.md)
6. 交付经过验证的结果
├─ 6.1 生成最终报告(report.md)
├─ 6.2 向用户展示主日志位置
├─ 6.3 用户审核与反馈
└─ 6.4 如果用户有 `/save-ob-chaos` 技能,执行并存档报告文件
AskUserQuestion Format Specification
All AskUserQuestion calls must follow this structure:
-
Re-ground (1-2 sentences)
- State the current project, branch, and task
- Explain the current progress
-
Simplify
- Explain the question in language a 16-year-old can understand
- Avoid internal jargon, use concrete examples
-
Recommend
RECOMMENDATION: 选择 [X] 因为 [一行理由]- Include
Completeness: X/10
-
Options
- A) B) C) format
- Show effort estimates
(human: ~X / CC: ~Y)
Example:
Current project: my-app | Branch: feature-search
You want to research the new features of React 19. In simple terms: you want to know what has changed in React 19 and whether it is worth upgrading.
RECOMMENDATION: Choose C (comprehensive investigation) because an upgrade decision requires a complete analysis of pros and cons
What do you want to achieve with this search?
A) Quick overview — 10-minute overview (Completeness: 5/10, human: ~30min / CC: ~3min)
B) Standard research — understand the main changes and migration costs (Completeness: 8/10, human: ~2h / CC: ~10min)
C) Comprehensive investigation — go deep into all new features, performance comparisons, and community feedback (Completeness: 10/10, human: ~1d / CC: ~20min)
Step 1: Requirements Clarification — Three-layer Ask Framework
Do not start searching directly. First clarify the research goals through structured questions.
1.1 Project Restatement (Re-ground)
Use AskUserQuestion:
Current project: {project-name} | Branch: {_BRANCH}
You asked me to conduct web search research. Let me first make sure I understand correctly:
Your original request was: "{user-original-request}"
I understand you want to research:{refined core topic}
Is that correct?
A) Yes, start the research
B) Not exactly — please clarify
1.2 Research Goal Classification
If the user confirms (chooses A), use AskUserQuestion:
What do you want to achieve with this search?
- Validate a hypothesis — you have a specific idea and need to find evidence for/against it
- Explore the unknown — You are unfamiliar with a certain field and need a systematic overview
- Comparative decision-making — You need to choose among several options (technology selection, solution comparison, etc.)
- Track the latest — Learn about the latest developments, versions, and trends in a certain field
- In-depth investigation — Comprehensively gather information to form a complete understanding
Mode mapping:
- Validate hypothesis / Comparative decision-making → Focus mode(Focused Mode)
- Explore the unknown / Track the latest → Scan mode(Scan Mode)
- In-depth investigation → Comprehensive mode(Comprehensive Mode)
1.3 Mode-specific questions (ask only one at a time)
Focus mode questions (Validate hypothesis / Comparative decision-making)
AskUserQuestion in sequence (intelligently skip questions already answered):
Q1: Specific hypothesis
What is your specific hypothesis? Please state the result you believe in one sentence.
Example:"I believe Next.js 14's App Router is more than 20% slower than the Pages Router"
Q2: Disproof criteria
What kind of evidence would make you change your mind? (Disproof criteria)
A) Official benchmark data is sufficient
B) Multiple independent sources must agree
C) Need to see actual code/cases
D) Other: ____
Q3: Scope boundaries
How broad is the research scope?
- Narrow — Focus on a specific version/scenario (e.g., only Next.js 14.2+)
- Medium — Include the related ecosystem (e.g., compare other frameworks)
- Broad — Historical evolution + future trends
Scan mode questions (Explore the unknown / Track the latest)
AskUserQuestion in sequence:
Q1: Background level
What is your current level of understanding of this topic?
- Complete beginner — Need explanations of basic concepts
- Some understanding — Know the main terminology, need to go deeper
- Very knowledgeable — Focus on the latest developments and edge cases
Q2: Information recency
Information recency requirements?
- Latest only — Information from 2024-2026
- Include history — Need the evolution trajectory
- No special requirements — Quality first
Q3: Output preference
What form of output would you like?
A) Executive summary (2-3 paragraphs of key conclusions)
B) Structured comparison table (pros and cons, applicable scenarios)
C) Detailed report (with citations, confidence assessment)
D) Raw findings list (you organize it yourself)
Comprehensive mode questions (In-depth investigation)
AskUserQuestion in sequence:
Q1: Research depth
How deep do you want the research to go?
- Quick overview — 1 round of search, completed within 30 minutes
- Standard research — 2-3 rounds of evaluator-optimizer loops
- Exhaustive research — Up to 4 rounds, until quality standards are met
Q2: Quality threshold
Confidence requirement?
- High — Multi-source cross-validation, contradictions must be resolved
- Medium — Primary sources being credible is sufficient
- Low — Quickly obtain directional information
Q3: Special angles
Do you need to focus on specific perspectives? (multiple selections allowed)
- Negative reviews / known limitations
- Chinese-language materials / domestic practice
- Academic papers / authoritative research
- GitHub Issues / real user feedback
- No special angle needed
1.4 Smart skip rules
Only ask questions whose answers are not yet clear.
| User's original request | Auto skip | Default setting |
|---|---|---|
| "I want to quickly get an overview" | Skip 1.2 goal classification | Scan mode |
| "Help me do in-depth research" | Skip 1.2 goal classification | Comprehensive mode |
| "Verify this idea" | Skip 1.2 goal classification | Focus mode |
| "Only need the latest information" | Skip Q2 in scan mode | latest-only |
| Provide a full technical comparison list | Comparison decision skips Q1 | Use the user's list |
1.5 Follow-up matrix (when the answer is not specific enough)
| User answer type | Follow-up question | Example follow-up wording |
|---|---|---|
| Vague/broad | Scope clarification | "This term will return a large number of results, and we need to narrow the scope. Which aspect/scenario specifically?" |
| No success criteria defined | Quality threshold | "What counts as 'finding the answer'? How high does the confidence need to be? (High=multi-source validation/Medium=mainstream sources/Low=preliminary understanding)" |
| No output preference | Output form | "Do you need a 3-sentence summary, a detailed report, or a raw list of findings?" |
| No time boundary | Timeliness requirement | "How new does the information need to be? Within 1 week/within 1 month/within 1 year/unlimited?" |
| No failure conditions | Stop condition | "Under what circumstances counts as 'not finding the answer'? When can the search stop?" |
1.6 Questioning quality principles
Anti-sycophancy rule — language that is prohibited
| ❌ Prohibited | ✅ Use instead |
|---|---|
| "This is an interesting question" | Directly state where the difficulty of the problem lies |
| "There are many ways to think about this" | Choose one and explain the reason |
| "You might consider..." | Say "should..." and explain why |
| "This can work" | Say "this can/cannot work because..." |
Follow-up discipline
- The first answer is usually a packaged version; the real answer comes after the second or third follow-up
- Push once, then push again — until the answer is specific and actionable
- Name common failure modes ("search term too broad", "no clear objective")
BAD vs GOOD examples
Scenario: The user says "search for LLM evaluation frameworks"
❌ BAD:
"OK, let me search for the latest information on LLM evaluation frameworks"
✅ GOOD:
"LLM evaluation frameworks are a big field, with hundreds of results. We need to narrow the scope:
- Do you want to compare mainstream frameworks (RAGAS/TruLens/DeepEval)?
- Or are you looking for evaluation for a specific task (such as code generation/mathematical reasoning)?
- Or do you want to understand evaluation methodologies (automated evaluation vs human evaluation)?
Which direction specifically?"
1.7 Escape hatch
If the user expresses impatience(such as "just search", "skip these questions"):
Understood, these questions do take an extra 2-3 minutes, but they determine search quality.
Quick mode: I will choose reasonable defaults based on your original request and start searching immediately. You can later ask to go deeper into a specific direction.
RECOMMENDATION: Choose B (continue asking questions) because 2-3 minutes of upfront clarification can avoid 20 minutes of ineffective searching
A) OK, quick mode (Completeness: 6/10)
B) No, continue asking until finished (Completeness: 10/10)
If the user expresses impatience a second time: respect the choice, go directly to step 2, and use the default configuration:
- Research mode: scan
- Quality threshold: medium
- Output format: executive summary
1.8 Output: Research Requirements Document (YAML)
Based on the user's answers, generate structured requirements:
研究目标: {提炼后的目标}
研究模式: {focused/scan/comprehensive}
背景水平: {beginner/intermediate/expert}
质量阈值: {high/medium/low}
时效要求: {latest-only/history-ok/none}
输出格式: {summary/comparison-table/detailed-report/raw-findings}
特殊角度: [negative, chinese, academic, github-issues, none]
反证标准: {string}
范围边界: {narrow/medium/wide}
Output the requirements document to:/tmp/search-web/{YYYY-MM-DD}/{task-name}/task.yaml
Directory structure preview:
/tmp/search-web/2026-04-05/powershell-bash-research/
├── task.yaml # 研究需求文档
├── master-log.md # 协调日志
├── report.md # 最终报告
├── turn-1/ # 第 1 轮
│ ├── search-agent-1-task.md
│ ├── search-agent-2-task.md
│ └── search-evaluator-turn-1.md
├── turn-2/ # 第 2 轮
│ ├── search-agent-3-task.md
│ ├── search-agent-4-task.md
│ ├── search-evaluator-turn-2.md
│ └── query-optimizer-turn-2.md
└── turn-3/ # 第 3 轮(最终轮无 optimizer)
├── search-agent-5-task.md
├── search-agent-6-task.md
└── search-evaluator-turn-3.md
This document will guide the query cluster construction in step 2 and be injected into the evaluator prompt.
Step 2: Evaluation Task and Query Construction
2.1 Determine the search mode
Based on the "Research Requirements Document" in step 1.8, determine $mode:
| Research mode + depth | $mode | Number of agents |
|---|---|---|
| focused/scan + quick overview | easy | 2 |
| comprehensive or standard research | medium | 3 |
| exhaustive research + high threshold | hard | 4 |
2.2 Choose the initial research direction
Choose a strategy based on the research mode:
Focused mode:
- Prioritize searching for evidence that supports/opposes the hypothesis
- Look for authoritative sources (official documentation, papers, benchmarks)
- Focus on contradictory views and counterevidence
Scan mode:
- Cast a wide net first to get an overview
- Identify key concepts and terms
- Find authoritative sources in the field
Comprehensive mode:
- Multiple angles in parallel: positive/negative/academic/practical
- Time dimension: historical evolution + latest progress
- Regional dimension: international mainstream + Chinese-language community
2.3 Build query clusters
Use the following techniques:
- Search techniques: side-by-side comparison, time limits, etc.
- Public benchmarks: named evaluation datasets/frameworks, etc.
- Specific sites: GitHub issues, HN, StackOverflow, etc.
- Academic reports: papers, arXiv, research blogs, etc.
- Multiple perspectives: such as negative bugs, limitations, complaints, etc.
- Multilingual: multilingual search, country- or region-specific search engines, etc.
Based on the "special angle" selected by the user, forcibly include the corresponding query:
negative→ "{topic} limitations problems issues"chinese→ "{topic} Chinese pitfalls practices"academic→ "{topic} paper arxiv research"github-issues→ "{topic} github issues bug"
2.4 Use the todos tool to create search tasks
TaskCreate:
subject: "网络搜索:{研究目标}"
description: |
研究目标:{提炼后的目标}
研究模式:{focused/scan/comprehensive}
质量阈值:{high/medium/low}
当前轮次:1/{max_rounds}
Step 3: Dispatch sub-agent searches
3.1 Create a search team
TeamCreate:
team_name: "search-web"
description: "网络搜索团队,用于并行执行多个搜索查询"
3.2 Launch parallel searches
Use the Agent tool together with subagent_type: general-purpose and team_name: search-web to launch parallel searches.
Search agent configuration:
name: search-agent-{n}
subagent_type: general-purpose
team_name: search-web
prompt: |
搜索:"{query}"
目标:{research_goal}
**输出文件:** 将结果写入 `/tmp/search-web/{YYYY-MM-DD}/{task-name}/turn-{round}/search-agent-{n}-task.md`
使用工具执行网络搜索。
**⚠️ 错误处理规则:**
- 如果遇到 429 Rate Limit 错误,**立即停止搜索**
- 返回已获取的部分结果 + 错误说明
- 不要无限重试或切换到其他工具继续搜索
对于每个相关结果:
1. 使用 WebFetch 获取页面内容
2. 提取与研究目标相关的关键发现
3. 记录:URL、标题、要点、相关度评分(1-5)
**写入文件格式:**
```markdown
# Search Agent {n} - Task Log
## 元数据
- **查询:** "{query}"
- **集群:** {cluster_name}
- **启动时间:** {timestamp}
- **状态:** {running/completed/error}
## 搜索结果
### 结果 1:{title}
- URL: {url}
- 相关度:{1-5}/5
- 关键发现:
- {要点}
- 引用:
- "{相关摘录}"
## 执行摘要
- 结果数量:{N}
- 平均相关度:{X}/5
- 遇到错误:{none/429/other}
If the results are insufficient, try variant queries. Focus on factual information and claims supported by data.
Return to the parent agent: Return only the file path and status summary (e.g., "Completed, 12 results, written to search-agent-1-task.md")
根据 `$mode`,**并行启动 2~4 个代理**(easy 2 个、medium 3 个、hard 4 个),每个代理使用不同的查询集群。
---
## 步骤 4:聚合结果
### 4.1 初始化会话目录
```bash
mkdir -p /tmp/search-web/{YYYY-MM-DD}/{task-name}
4.2 Create a coordination log file
The main agent only records the coordination log, in the format of single-line timestamped records:
File path: /tmp/search-web/{YYYY-MM-DD}/{task-name}/master-log.md
Format:
[14:30:00] [INIT] 搜索任务初始化,研究目标:{research_goal}
[14:30:05] [MKDIR] 创建目录 turn-1/
[14:30:05] [AGENT_START] search-agent-1 启动,查询:{query_1},集群:comparison,输出:turn-1/search-agent-1-task.md
[14:30:05] [AGENT_START] search-agent-2 启动,查询:{query_2},集群:limitations,输出:turn-1/search-agent-2-task.md
[14:30:50] [AGENT_DONE] search-agent-1 完成,结果数:12,耗时:45s
[14:30:52] [AGENT_DONE] search-agent-2 完成,结果数:8,耗时:38s
[14:31:00] [EVAL_START] 启动 search-evaluator,第 1 轮评估,输出:turn-1/search-evaluator-turn-1.md
[14:33:30] [EVAL_DONE] 评估完成,总分:3.2/5,建议:CONTINUE_SEARCH
[14:33:35] [ROUND_END] 第 1 轮结束,启动第 2 轮
[14:33:36] [MKDIR] 创建目录 turn-2/
[14:33:37] [OPTIM_START] 启动 query-optimizer,输出:turn-2/query-optimizer-turn-2.md
[14:35:00] [AGENT_START] search-agent-3 启动,输出:turn-2/search-agent-3-task.md
...
[14:45:00] [COMPLETE] 搜索完成,总轮次:3,最终评分:4.5/5
Coordination log principles:
- Record only event type, timestamp, and key metadata
- Do not record detailed search results (recorded by the sub-agents' own files)
- Facilitates quick tracing of the execution flow and locating issues
Step 5: Evaluator - optimizer loop
Use two sub-agents to form a feedback loop: the evaluator evaluates the search results, and the optimizer optimizes the search based on the evaluation results.
Evaluator tone: challenging and destructive
5.1 Evaluator agent
name: search-evaluator
subagent_type: general-purpose
team_name: search-web
timeout: 180000
prompt: |
你是研究质量评估器。根据我们的研究目标和质量标准评估以下搜索结果。
研究目标:{goal}
质量阈值:{high/medium/low}(来自步骤 1.8)
研究模式:{focused/scan/comprehensive}
**读取来源:** 从当前轮次目录读取所有搜索代理结果
- `/tmp/search-web/{YYYY-MM-DD}/{task-name}/turn-{round}/search-agent-*.md`
**输出文件:** 将评估报告写入 `/tmp/search-web/{YYYY-MM-DD}/{task-name}/turn-{round}/search-evaluator-turn-{round}.md`
从以下维度评估(每项 1-5 分):
1. **覆盖度** - 结果是否从多个角度回答了核心问题?
2. **来源质量** - 来源是否权威可信?
3. **一致性** - 来源在关键事实上是否一致?
4. **具体性** - 论断是否有数据/案例支持?
5. **时效性** - 信息是否最新且相关?
特殊规则(来自步骤 1.8):
- 若用户要求"仅最新":2024-2026 信息正常评分,2023 年及以前降 2 分
- 若用户要求"包含历史":不扣分,但需标注时间
**写入文件格式:**
```markdown
# Search Evaluator - Turn {round} Report
## 元数据
- **评估轮次:** {round}
- **评估时间:** {timestamp}
- **研究目标:** {goal}
- **质量阈值:** {high/medium/low}
## 质量评估
| 维度 | 评分 | 依据 |
|------|------|------|
| 覆盖度 | {1-5} | {简要说明} |
| 来源质量 | {1-5} | {简要说明} |
| 一致性 | {1-5} | {简要说明} |
| 具体性 | {1-5} | {简要说明} |
| 时效性 | {1-5} | {简要说明} |
**总分**:{average}/5
## 缺口分析
缺失或需要改进的内容:
1. {具体缺口 1}
2. {具体缺口 2}
## 改进建议
为解决缺口,搜索:
1. {具体后续查询 1}
2. {具体后续查询 2}
## 完成评估
**退出条件(必须全部满足):**
- [ ] 覆盖度 > 4.5(即 5/5)
- [ ] 来源质量 > 4.5(即 5/5)
- [ ] 一致性 > 4.5(即 5/5)
- [ ] 具体性 > 4.5(即 5/5)
- [ ] 时效性 > 4.5(即 5/5)
**建议**:{CONTINUE_SEARCH / COMPLETE}
> ⚠️ **硬性标准**:只有所有维度评分均为 5/5 时才能标记为 COMPLETE。任一维度 ≤4 必须继续搜索。
Return to the parent agent: Return only the file path and key conclusions (e.g., "Score 3.2/5, recommend CONTINUE, written to turn-1/search-evaluator-turn-1.md")
### 5.2 决策点
**硬性退出标准:所有维度评分 > 4.5(即 5/5)**
基于评估器输出:
**如果所有维度 = 5/5(COMPLETE):**
- 进入步骤 6
- 在主日志中记录最终评估
**如果任一维度 ≤ 4(CONTINUE_SEARCH):**
- 提取改进建议
- 针对 ≤4 的维度优先构建新查询
- 返回步骤 3(最多 4 轮)
**注意**:不允许"综合判断"或"总体得分"绕过此标准。任一维度不达标就必须继续。
### 5.3 优化器代理(复杂情况可选)
```yaml
name: query-optimizer
subagent_type: general-purpose
team_name: search-web
timeout: 180000
prompt: |
你是搜索查询优化器。基于评估器反馈,优化搜索策略。
原始目标:{goal}
**读取来源:** 从当前轮次目录读取评估报告
- `/tmp/search-web/{YYYY-MM-DD}/{task-name}/turn-{round}/search-evaluator-turn-{round}.md`
**输出文件:** 将优化建议写入 `/tmp/search-web/{YYYY-MM-DD}/{task-name}/turn-{round}/query-optimizer-turn-{round}.md`
基于评估器识别的缺口,生成 2-3 个新的查询集群。
避免重复先前的查询。
**写入文件格式:**
```markdown
# Query Optimizer - Turn {round} Report
## 元数据
- **优化轮次:** {round}
- **优化时间:** {timestamp}
- **参考评估:** evaluator-turn-{previous_round}.md
## 评估器反馈摘要
- 主要缺口:{gap_summary}
- 关键评分:{scores}
## 优化的查询集群
### 集群 1:{缺口名称 A}
- 主要:"{优化后的查询}"
- 变体:
- "{变体 1}"
- "{变体 2}"
- 目标站点:{特定站点}
- 理由:{为什么这能解决缺口}
### 集群 2:{缺口名称 B}
...
## 预期改进
- 覆盖度提升:{X} → {Y}
- 具体性提升:通过搜索 {specific_query}
Return to the parent agent: Return only the file path and number of clusters (e.g., "Generated 3 clusters, written to turn-2/query-optimizer-turn-2.md")
---
## 步骤 6:交付经过验证的结果
### 6.1 最终摘要
向用户返回精简的研究摘要,然后撰写详细报告。
报告写入 `/tmp/search-web/{YYYY-MM-DD}/{task-name}/report.md`:
```markdown
# 研究报告:{task-name}
## 执行摘要
{回答核心问题的 2-3 段文字}
## 关键发现(含置信度)
| 发现 | 证据 | 置信度 |
|------|------|--------|
| {要点 1} | {来源} | 高/中/低 |
| {要点 2} | {来源} | 高/中/低 |
## 质量评估
- 覆盖度:{score}/5 - {简要说明}
- 来源质量:{score}/5 - {简要说明}
- 一致性:{score}/5 - {简要说明}
## 局限性
{仍不确定或需要进一步调查的内容}
## 来源
{带 URL 的关键引用}
6.2 Main log location
Point the user to the detailed logs:
Full search session log:
/tmp/search-web/{YYYY-MM-DD}/{task-name}/master-log.md
6.3 User review
Ask:
- "Research complete, quality score {X}/5. Is it sufficient?"
- "Would you like me to investigate any specific findings in depth?"
Multi-round protocol (with evaluator feedback)
Round 1: Broad exploration
- Cast a wide net with general queries
- The evaluator identifies key gaps
Round 2: Targeted investigation
- Address specific gaps from the Round 1 evaluation
- Search for contradictory evidence
Round 3+: Deep dive
- Focus on unresolved issues from the evaluator
- Search primary sources to fill gaps
Stop conditions:
- All dimension scores = 5/5(hard criterion, no comprehensive judgment allowed)
- Reach the round limit (default: 4)
- User forces a stop
Quality standards reference
Score calibration
| Score | Coverage | Source quality | Consistency | Specificity | Timeliness |
|---|---|---|---|---|---|
| 5 | All key angles deeply covered | Multiple authoritative sources | Strong consensus | Grounded in specific data | Latest and dated |
| 4 | Most angles covered | Mainly authoritative | Minor discrepancies | Mostly specific | Mostly latest |
| 3 | Gaps but acceptable | Mixed quality | Partially contradictory | Mix of specific and general | Partially outdated |
| 2 | Major gaps | Mainly low quality | Major contradictions | Mainly general | Clearly outdated |
| 1 | Missing key perspectives | Unverified sources | Completely contradictory | Vague assertions | Severely outdated |
Best practices
- The evaluator is mandatory — never skip quality assessment
- Hard exit criteria — all dimensions must reach 5/5, no "overall good" composite judgment allowed
- Feedback must be specific — vague "search a bit more" is not helpful
- Track quality trends — if scores are not improving, reconsider the approach
- Record evaluator output — each evaluation is valuable for transparency
- Early clarification determines later quality — do not skip the questions in Step 1
- Probe until specific — comfort means not deep enough
Timeliness Search Guide
Current date: 2026-04-05
Timeliness-sensitive topics
Add timeliness filters or time-bounded queries for the following:
- Technology versions(software releases, framework updates)
- Market data(prices, trends, statistics)
- Current events(news, regulations, policies)
- Research papers(latest publications, SOTA results)
Timeliness query patterns
# 最新版本
"{topic} latest version 2025 2026"
"{topic} release notes 2026"
# 当前最佳实践
"{topic} best practices 2025"
"{topic} guide 2026"
# 破坏性变更
"{topic} breaking changes 2025 2026"
"{topic} migration guide"
# 研究
"{topic} paper 2025 2026"
"{topic} benchmark 2025"
Evaluator timeliness adjustments
When evaluating results, consider the search date:
- Information from 2025-2026: current (normal weight)
- Information from 2024: may be outdated for fast-moving fields
- Information from 2023 or earlier: may be outdated for technical topics
It is currently 2026, and you need to adjust accordinglytimelinessscore, and flag key information that may be outdated.
Example session
User request: "Search for evidence that LLM agents perform worse on PowerShell than Bash"
Step 1.1 (project restatement):
- The user confirms the research objective: verify the performance difference of LLM agents on PowerShell vs Bash
Step 1.2 (objective classification):
- The user selects "verify hypothesis" → enter focused mode
Step 1.3 (focused mode):
- Q1 hypothesis: "LLM agents have a higher error rate on PowerShell"
- Q2 Falsification standard: multiple independent benchmarks are needed
- Q3 Scope: narrow (code generation tasks only)
Step 1.8 (requirements document):
研究目标: 验证 LLM 代理在 PowerShell vs Bash 的表现差异
研究模式: focused
质量阈值: high
时效要求: latest-only
输出格式: detailed-report
特殊角度: [github-issues]
Round 1:
- Query: general comparison, benchmark search
- Evaluator: "Coverage 3/5 - needs more specific benchmarks; Specificity 2/5 - needs actual numbers"
- Recommendation: CONTINUE
Round 2:
- Query: SWE-bench results, pass@k metrics, GitHub issue analysis
- Evaluator: "Coverage 4/5; Specificity 4/5 - found good data; Consistency 3/5 - some contradictory claims"
- Recommendation: CONTINUE
Round 3:
- Query: focus on contradictory claims, primary sources
- Evaluator: "All dimensions = 5/5. Meets hard exit criteria."
- Recommendation: COMPLETE
Deliverable: validation summary with confidence scores and quality assessment