LLMSkillHub
mirroredCopied from oxbshw/LLM-Agents-Ecosystem-Handbook under MIT. oxbshw has not claimed this page, and is not affiliated with LLM SkillHub.

oxbshw / research-summarizer

Use when the user asks for a sourced briefing on a topic that spans multiple web sources and requires citations.

0
SKILL.md
2.1 KB

Research Summarizer

When to use

  • User asks "research X" / "what's the state of Y" / "summarize the literature on Z"
  • The answer requires multiple sources, not one
  • The user expects citations

When NOT to use

  • Single-source summarization (use a plain summarize tool)
  • Code search (use repo-auditor or grep)
  • Opinion / recommendation requests (this skill produces neutral briefings)

Inputs

NameTypeRequiredNotes
topicstringyesthe question to research
depth"shallow" | "deep"nodefault "shallow" (5–7 sources); "deep" → 10–15
audiencestringnoshapes vocabulary level (e.g., "executive", "engineer")

Outputs

A Markdown file briefing.md following the structure in references/report-template.md.

Workflow

  1. Plan: produce 3–7 sub-questions covering breadth and depth
  2. Search + fetch: for each sub-question, run web_search then fetch_url for top 2–3 hits
  3. Cluster: group findings by claim; load references/clustering-rules.md
  4. Draft: fill references/report-template.md — every claim needs a [n] citation
  5. Validate: every citation resolves; no domain dominates (≥ 3 distinct domains for shallow, ≥ 5 for deep)
  6. Self-review against success criteria below

References

  • [`references/report-template.md`](references/report-template.md) — the output structure
  • (add clustering-rules.md, domain-allowlist.md as the skill matures)

Success criteria

  • ≥ 5 sources for shallow, ≥ 10 for deep
  • ≥ 3 distinct domains
  • 0 unsourced claims (or all flagged (unsourced))
  • Open Questions section populated

Failure modes

  • Fewer sources available than required → return partial result, flag at top
  • All sources from one domain → reject, retry plan with broader scope
  • Network error → fail loudly, don't fabricate

Examples

  • See references/examples/ (TBD) for one shallow and one deep run