菜单
Use when a task involves web search, current information, source verification, citations, or entity/person/paper/project disambiguation.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
| name | agentic-search |
| description | Use when a task involves web search, current information, source verification, citations, or entity/person/paper/project disambiguation. |
Use this skill when search results can become evidence for an answer. It governs agent behavior around web search, source authority, current information, and entity disambiguation. It does not implement a search backend, RAG system, or provider-specific router behavior.
Search results are leads, not conclusions. Convert the user's request into checkable claims, inspect source authority, disambiguate entities, and downgrade the answer when evidence is weak or conflicting.
Use this skill when the task involves any of these:
If the task is computational search, parameter search, benchmark search, or
optimization search, use brain:think-before-you-calculate instead. If both
external source judgment and computation are involved, use both skills and keep
their boundaries separate.
Identify what is being searched before running or interpreting search:
Object Type: fact | current status | person | paper | project | repository | organization | product | standard | broad topic
Target Object:
Known Discriminators:
Missing Discriminators:
For ambiguous entities, collect discriminators such as affiliation, field, date range, title words, coauthors, location, repository owner, organization, publication venue, or user-provided context.
Do not search a vague topic and report whatever appears. Convert the request into one or more claims:
Claim:
Evidence Needed:
Source Class Needed:
Failure Mode:
Examples:
Claim: This search result refers to the same person as the user-provided context.
Evidence Needed: Matching affiliation, field, or first-party page.
Source Class Needed: official page, institutional page, author profile, repository, or primary publication.
Failure Mode: Same-name person from another institution or field.
Claim: This paper is the requested work, not a similarly titled paper.
Evidence Needed: Matching title, authors, venue or arXiv/DOI, and method context.
Source Class Needed: paper page, PDF, arXiv/OpenReview/publisher page, author page, or repository.
Failure Mode: Similar title phrase with different authors or topic.
Prefer sources in this order:
Do not treat search rank, snippets, or model summaries as authority by themselves.
For entity-sensitive tasks, use targeted query variants rather than one generic query:
"<name>" "<affiliation>"
"<name>" "<field keyword>"
"<name>" "<coauthor or collaborator>"
"<paper title>" "<author>"
"<paper title>" "<venue or year>"
"<repository>" "<owner or organization>"
"<project acronym>" "<domain keyword>"
Actively look for near misses:
Use one of these statuses before answering:
| Status | Meaning |
|---|---|
confirmed | Authoritative or primary sources directly support the claim. |
likely | Good sources support the claim, but one discriminator or primary source is missing. |
unresolved | Sources are insufficient or ambiguous. |
conflicting | Credible sources disagree or point to different entities. |
When the user is the subject or directly owns the private fact, user-provided ground truth is authoritative for that identity or first-party context. Search evidence must be reconciled with it, not placed above it.
The final answer should state:
Use concise language for simple searches. Use a compact evidence table for entity-sensitive, disputed, or multi-source answers.
Apply this frame internally when search output affects a claim:
Whole Object: The external fact or entity identity being claimed.
State: Search results, source pages, user-provided context, and current date.
Representation: Query strings, snippets, source URLs, page text, citations.
Control Chain: User request -> agent query choice -> search tool/provider -> source inspection -> answer.
Bypass Path: Accepting a search snippet or first result without source inspection.
Failure Path: Wrong entity, stale source, source conflict, or overconfident synthesis.
Responsibility Owner: The agent owns evidence boundaries; the user owns private or first-party ground truth they explicitly provide.
Search and cite current sources. Prefer official or primary sources. If the source date is missing or stale, say so.
Do not assume the first matching name is the requested person. Match at least one
strong discriminator such as affiliation, field, coauthor, location, official
profile, or user-provided context. If none match, answer unresolved or
conflicting.
Prefer DOI, arXiv, OpenReview, publisher, conference, author page, PDF, or repository links. Title similarity is insufficient when authors, venue, year, or topic do not match.
Prefer official documentation, changelogs, release notes, or source code over blog posts and snippets. If the user asks about OpenAI products, follow the official OpenAI documentation requirement.
llm_routerThis skill does not change llm_router. Keep responsibilities separate:
| Layer | Responsibility |
|---|---|
llm_router | Protocol compatibility, provider payload translation, Responses state correctness. |
brain:agentic-search | Agent search strategy, source authority, entity disambiguation, conflict handling. |
| User | Private ground truth, task acceptance, and domain context only they can know. |
Better retrieval does not remove the need for claim-level verification.
For disputed or entity-sensitive tasks, prefer this compact structure:
Status: confirmed | likely | unresolved | conflicting
Claim:
Evidence:
Disambiguators Checked:
Remaining Gap:
Answer:
For ordinary current-info tasks, a short answer with citations and a note about uncertainty is enough.
Search should reduce uncertainty, not launder uncertainty into confidence. Name the object, verify the claim, disambiguate near misses, and mark the boundary of what the sources actually support.
Use when installing, reinstalling, auditing, or preparing Brain-managed Superpowers, Spec Kit, and OpenSpec workflow arsenals under a project-local `.brain/vendor` directory so Brain can inspect and route them without exposing their raw Codex skills/prompts at the project root.
Use when the user asks to check, add, or manage macOS Calendar events.
Use when the user asks to send an email via the macOS Mail app.
Use when the user asks to send a text message or iMessage via the macOS Messages app.
Use when the task involves `exec_command` escalation, `prefix_rule`, sandbox permissions, writable roots, or Codex platform-specific operational patterns.
Use when a task asks to run calculations, simulations, model training, computational searches, benchmarks, optimizations, workflows, agents, pipelines, or other tool-heavy execution before the scientific question and interpretation boundary are clear.