| name | x-ipe-tool-web-search |
| description | Delegate focused web research to the current coding agent's web capability, then return structured findings and citations. Use when a skill needs industry best practices, standards, competitor patterns, or library/API guidance from the web. Triggers on requests like "web search", "research topic", "find best practices", "search the web". |
Web Search Tool
Purpose
AI Agents follow this skill to perform focused external research by:
- Determining whether a question is appropriate for web research
- Delegating search and retrieval to the current agent's web capability
- Synthesizing findings into reusable principles with citations
Important Notes
BLOCKING: Use this skill ONLY for general knowledge topics such as industry standards, best practices, regulations, competitor patterns, or public API/library guidance.
BLOCKING: Do NOT use this skill for project-internal questions about repository code, local files, workflow state, or unpublished requirements.
CRITICAL: Prefer the current coding agent's native web capability first. Examples include built-in web search, internet-enabled fetch tools, or dedicated browsing/fetch tools exposed by the agent runtime.
CRITICAL: Keep research bounded. Default to 3 useful sources maximum unless the caller explicitly needs more.
About
This tool skill standardizes how X-IPE skills ask an agent to perform public web research. It does not invent a separate crawler or search engine. Instead, it tells the executing agent to use whatever web capability the CLI/runtime already provides, then normalize the result into a compact, reusable structure.
Key Concepts:
- General Knowledge Request — A topic that can be answered from public web sources without needing project-private context
- Agent Web Capability — Any built-in web search, browsing, fetch, or retrieval ability available to the current CLI agent
- Seed URLs — Caller-provided links that should be fetched before broader search
- Recommended Principles — Actionable, caller-friendly takeaways extracted from the research findings
When to Use
triggers:
- "web search"
- "research topic"
- "find best practices"
- "search the web"
- "research common principles"
- "research industry standards"
not_for:
- "Project-internal code or file questions — use repo search or explore instead"
- "Autonomous conflict or routing resolution — use x-ipe-assistant-user-representative-Engineer"
Input Parameters
input:
operation: "research_topic"
research_request:
topic: "short topic name"
goal: "what the caller needs to learn"
questions: ["specific question 1", "specific question 2"]
seed_urls: ["https://..."]
preferred_source_types: ["official-docs", "standards", "vendor-docs", "reputable-guides"]
max_sources: 3
Input Initialization
<input_init>
<field name="operation" source="Caller specifies the operation">
<validation>MUST equal `research_topic`.</validation>
</field>
<field name="research_request.topic" source="Caller provides the public topic to research">
<validation>MUST be non-empty and MUST NOT reference project-private files or code.</validation>
</field>
<field name="research_request.goal" source="Caller provides why the research is needed">
<validation>MUST be non-empty.</validation>
</field>
<field name="research_request.questions">
<steps>
1. IF caller provided explicit questions → use them.
2. ELSE derive 2-3 focused questions from topic + goal.
</steps>
</field>
<field name="research_request.seed_urls">
<steps>
1. IF caller provided seed URLs → fetch them first.
2. ELSE use the agent's native web capability to discover candidate sources.
</steps>
</field>
<field name="research_request.max_sources">
<steps>
1. IF caller provided a value → use it.
2. ELSE default to 3.
3. Cap at 5 to keep research bounded.
</steps>
</field>
</input_init>
Definition of Ready
<definition_of_ready>
<checkpoint required="true">
<name>General topic identified</name>
<verification>research_request.topic is public/general knowledge, not project-internal</verification>
</checkpoint>
<checkpoint required="true">
<name>Research goal provided</name>
<verification>research_request.goal explains what the caller needs from the research</verification>
</checkpoint>
<checkpoint required="true">
<name>Web capability available</name>
<verification>The executing agent has at least one web-capable search/fetch mechanism available</verification>
</checkpoint>
</definition_of_ready>
Operations
Operation: research_topic
When: A calling skill needs external research on a public topic such as standards, best practices, competitor patterns, or public API/library usage.
<operation name="research_topic">
<action>
1. Validate the request:
- Confirm topic and goal are present.
- Reject the request if it is project-specific or references private repo state.
2. Build the research plan:
- Normalize the questions list.
- If seed URLs were provided, use them first.
- Else formulate 1-2 focused search prompts from topic + goal.
- Prefer the current agent's native web capability to discover authoritative sources.
3. Gather sources:
- Fetch or read the selected sources with the available web capability.
- Prioritize official docs, standards bodies, regulators, major vendor docs, and reputable engineering references.
- Exclude duplicates, irrelevant pages, and low-signal summaries.
- Keep at most {max_sources} useful sources.
4. Extract findings:
- Capture per-source notes: title, URL, authority type, and the findings relevant to the caller's questions.
- Note disagreements or uncertainty across sources.
5. Synthesize the result:
- Consolidate recurring findings into key_findings.
- Convert those findings into recommended_principles the caller can apply immediately.
- Return concise citations so downstream skills can reference them.
</action>
<constraints>
- BLOCKING: Fail with `WEB_SEARCH_NOT_APPROPRIATE` if the request is project-specific
- BLOCKING: Fail with `WEB_CAPABILITY_UNAVAILABLE` if the executing agent cannot reach any web capability
- CRITICAL: Do not exceed 5 fetched sources in one call
- CRITICAL: Prefer authoritative sources even if they are less concise than blog summaries
</constraints>
<output>operation_output with findings, recommended principles, and source citations</output>
</operation>
Output Result
operation_output:
success: true | false
result:
topic: "researched topic"
goal: "caller goal"
key_findings:
- summary: "finding"
supports_questions: ["question 1"]
recommended_principles:
- "actionable principle"
sources:
- title: "source title"
url: "https://..."
authority: "official-docs | standards | vendor-docs | reputable-guide"
notes: "why this source was used"
uncertainties:
- "open question or disagreement"
errors: []
Definition of Done
<definition_of_done>
<checkpoint required="true">
<name>Request Validated</name>
<verification>General/public topic confirmed and request rejected if project-specific</verification>
</checkpoint>
<checkpoint required="true">
<name>Bounded Source Set Selected</name>
<verification>Candidate sources were prioritized and limited to configured maximum</verification>
</checkpoint>
<checkpoint required="true">
<name>Structured Findings Returned</name>
<verification>operation_output includes key_findings, recommended_principles, and sources or a clear failure reason</verification>
</checkpoint>
</definition_of_done>
Error Handling
| Error | Cause | Resolution |
|---|
WEB_SEARCH_NOT_APPROPRIATE | Request is project-specific or private | Use repository exploration tools instead of web search |
WEB_CAPABILITY_UNAVAILABLE | Current agent/CLI has no usable web capability | Fall back to internal knowledge only or run on a web-capable agent |
NO_USEFUL_SOURCES | Search produced no relevant authoritative sources | Narrow the topic, add seed URLs, or retry with sharper questions |
Templates
| File | Purpose |
|---|
templates/research-summary.md | Optional markdown structure for callers that want to embed research findings verbatim |
Examples
See .github/skills/x-ipe-tool-web-search/references/examples.md for usage examples.
