deep-research

v2026.09.24

Conduct deep, source-grounded research with freshness checks and citations, then synthesize an actionable answer. Use for time-sensitive, niche, or high-stakes topics; when the user asks to “zweryfikuj”, “zrób research”, or when uncertainty is high.

GitHub
安装命令
npx skhub add janjaszczak/deep-research
Markdown
SKILL.md

deep-research

Activation gate (anti-noise)

Run only if at least one is true:

  • User explicitly requests research/verification or provides external links.
  • Topic is time-sensitive (prices, laws, tooling versions, security, vendor docs).
  • You detect >10% risk of outdated knowledge. Otherwise: answer normally (no research ritual).

Research protocol

1) Frame

  • Restate the question as a falsifiable query.
  • Define “freshness window” (e.g., last 3–6 months) unless user asks historical.
  • Define acceptance criteria for sources (primary docs preferred).

2) Collect sources

  • Prefer: official docs, primary specs, reputable engineering blogs, vendor changelogs.
  • For each key claim, capture at least one authoritative source.
  • If sources disagree: represent both and explain implications.

Tool hierarchy (cost-aware):

  1. Named library/framework API docs (version-aware): context7 MCP + skill mcp-context7-docs — not general web search.
  2. Default: searxng MCP (self-hosted, free) or duckduckgo MCP (free) for routine lookups and freshness checks — the agent synthesizes results itself.
  3. Escalate to perplexity MCP only when the user explicitly asks for "deep research"/"zweryfikuj dokładnie", or the query is high-stakes and a single-call synthesized answer with citations justifies the per-request cost.
  4. Cursor-bundled parallel-deep-research/parallel-web-search skills only when the user explicitly says "deep"/"exhaustive"/"comprehensive report" (see that skill's own activation gate) — these are the most expensive option.

3) Extract & normalize

  • Extract only what is needed to answer.
  • Keep notes in a compact “evidence table” (claim → source → date).

4) Synthesize

Produce one of:

  • Research brief (short, decision-oriented)
  • Decision memo (trade-offs + recommendation)
  • Implementation notes (steps + verification commands)

5) Verification hooks (optional)

If the environment allows:

  • Provide 1–3 concrete commands/tests to verify claims locally.
  • For Cursor, prefer commands that are fast and deterministic.

Optional: store memo to MCP memory

If the user says “zapisz”:

  • Store: question, key findings, chosen decision, links, and date.

Output contract

  • Lead with recommendation.
  • Then evidence-backed reasoning (citations).
  • End with “how to verify quickly”.

Failure handling

If web/tools not available:

  • Mark claims as UNCERTAIN.
  • Provide the shortest verification path (exact query + official doc location).
发现
标签

此技能尚未发布标签。

版本
最新版本元数据

版本

v2026.09.24

发布时间

Sep 24, 2026

分类

未分类

许可证

未指定

源路径

skills/deep-research

默认分支

main

最新提交

cb70ff7

Tree SHA

dc4c45a