AGENT SKILL

research_current_info

Use when the user asks for up-to-date or current information, to confirm something is still accurate, or when a task depends on facts that may have changed since training (library versions, APIs, pricing, releases, news, current best practices). Runs a governed web search with a curated domain allow/deny policy and cited, recency-checked results.

License 许可
Apache-2.0
最近更新
2026-09-13
Stars
★ 0
Forks
0

01 前言 Front Matter

name
research_current_info
description
Use when the user asks for up-to-date or current information, to confirm something is still accurate, or when a task depends on facts that may have changed since training (library versions, APIs, pricing, releases, news, current best practices). Runs a governed web search with a curated domain allow/deny policy and cited, recency-checked results.

03 SKILL.md 正文

Skill: research_current_info

Purpose

Ground answers in current, verifiable web sources instead of stale memory, applying a curated domain policy (prefer trustworthy/primary sources, avoid noisy ones) and always citing what was found with its retrieval date.

When To Use

  • The user says "latest", "current", "today", "still true?", "has this changed?".
  • A task depends on volatile facts: library/tool versions, API signatures, prices, release dates, security advisories, current best practices.
  • Skip for timeless/internal knowledge — do not search when stable reasoning or the repo's own code already answers it.

Required Input

  • The question or claim to verify (topic + what "current" means here).
  • Optional: time window (e.g. "last 12 months"), and per-query domains to prefer/avoid on top of the policy below.

Domain Policy (EDIT THIS — it is your governance)

This section is the curated allow/deny list. Keep it under version control and tune it to your stack. The engine propagates it to every tool on make sync-skills.

Prefer (allow-list — primary / authoritative):

  • docs.python.org, peps.python.org
  • docs.astral.sh (uv, ruff), official docs of the libraries in pyproject.toml
  • github.com (source, releases, issues), pypi.org
  • developer.mozilla.org, official cloud/provider docs, arxiv.org

Avoid (deny-list — noisy / low-signal for authoritative detail):

  • SEO content farms and auto-generated mirror sites.
  • Examples often deprioritized for API correctness (keep or drop as you see fit): w3schools.com, geeksforgeeks.org — fine for a quick snippet, weak as a source of truth.

Rule of precedence: primary/official source > reputable secondary (well-known blogs, Stack Overflow) > everything else. If a query genuinely needs a domain not on the allow-list, use it but flag that it is outside the policy.

Execution Rules

  1. Use the runtime's web search tool. On Claude Code, pass the policy to WebSearch via allowed_domains / blocked_domains; on runtimes without domain filters, apply the policy by filtering/ranking results manually.
  2. Cross-check every load-bearing claim against at least two independent sources.
  3. Prefer primary/official sources; treat blogs and forums as secondary corroboration.
  4. Cite every external claim: title — URL — retrieval date. Never fabricate a URL.
  5. Distinguish verified facts from speculation; call out when sources disagree or when information may already be stale.
  6. If search returns nothing usable within the policy, say so explicitly rather than guessing — then ask whether to widen the allow-list.
  7. Keep all artifacts in English.

Output Format

  • A concise answer grounded in the sources.
  • Sources: a list of title — URL — date.
  • Confidence: high/medium/low with any caveats, conflicts, or staleness notes.

04 同仓库更多 Skills