pipeline-research

Full research pipeline: auto-researcher → firecrawl-agent → graphify → notebooklm-memory. Turns a research question or topic into a structured knowledge asset — multi-source synthesis, crawled source

SKILL.md

GitHub

Pipeline: Research

Full research pipeline. Five stages, sequential. Each builds on the previous.

Anti-redundancy: This pipeline calls auto-researcher, firecrawl-agent, graphify, and notebooklm-memory internally. Do NOT add them as separate steps after running this pipeline.


Inputs

Required from user or calling context:

  • Topic / question: what to research (required)
  • Depth: quick (30 min), standard (2 hours), deep (half-day). Default: standard
  • Output path: where to save research files. Default: {project}/memory/research/{topic-slug}/
  • Sources to crawl: optional list of URLs to deep-crawl. If none given, auto-researcher selects.

Stage 1 — Multi-Source Research (auto-researcher)

Invoke /auto-researcher:

  • Topic: {topic}
  • Depth: {depth}
  • Output confidence tiers: CONFIRMED / SUPPORTED / INFERRED / SPECULATIVE

Collect:

  • Key facts with confidence tier
  • Source URLs (categorized: primary, secondary, opinion)
  • Knowledge gaps identified
  • Recommended URLs for deep crawl

Write synthesis to {output-path}/synthesis.md.


Stage 2 — Deep Crawl High-Value Sources (firecrawl-agent)

From Stage 1, select top 3-5 source URLs:

  • Prioritize: official docs, primary sources, detailed technical writeups
  • Skip: aggregators, news reposts, thin content

Invoke /firecrawl-agent on each selected URL:

  • Mode: structured extraction
  • Extract: key claims, data points, quotes, links to deeper pages
  • Depth: 1 level deep per URL (avoid crawling entire sites)

Write extracted content to {output-path}/sources/ (one file per URL).


Stage 3 — Build Knowledge Graph (graphify)

Invoke /graphify:

  • Input: synthesis.md + all files in sources/
  • Extract: entities, relationships, concepts, contradictions
  • Output: knowledge graph at {output-path}/graph.json

Identify and flag:

  • Contradictions between sources
  • High-confidence vs speculative claims
  • Key entities and their relationships

Write graph summary to {output-path}/graph-summary.md.


Stage 4 — Sync to NotebookLM (notebooklm-memory)

Invoke /notebooklm-memory:

  • Topic: {topic}
  • Sources: synthesis.md, graph-summary.md, and top 3 source files
  • Action: create notebook if not exists, add sources, write index note

This creates a persistent, queryable NotebookLM notebook for long-term retrieval.


Stage 5 — Produce Research Report

Write {output-path}/research-report.md:

# Research Report — {topic}

**Date:** {date}
**Depth:** {depth}
**Sources crawled:** {N}

## Key Findings (CONFIRMED)
{bullet list of high-confidence findings}

## Supporting Evidence (SUPPORTED)
{bullet list of findings backed by multiple sources but not definitive}

## Inferred / Speculative
{bullet list with explicit confidence labels}

## Knowledge Gaps
{what remains unknown or unresolvable from available sources}

## Source Quality
{brief assessment of source diversity and reliability}

## Files
- Synthesis: {output-path}/synthesis.md
- Sources: {output-path}/sources/
- Knowledge graph: {output-path}/graph.json
- NotebookLM: {notebook-url or "synced — search via /notebooklm-memory"}

Print summary:

PIPELINE-RESEARCH COMPLETE
Topic: {topic}
Sources: {N crawled}
Findings: {N confirmed, M supported, K inferred}
Report: {output-path}/research-report.md
NotebookLM: {synced / failed}

Quality Gates

  • Stage 1 must produce at least 3 CONFIRMED findings before proceeding to Stage 2. If fewer than 3: widen search scope or flag insufficient public information.
  • Stage 2 must successfully crawl at least 1 source before proceeding. If all crawls fail: proceed with Stage 1 synthesis only, note in report.
  • Stage 3 graph must have at least 5 nodes. If fewer: note sparse graph in report.
  • Stage 4 failure (NotebookLM unavailable): skip and note in report. Do not block.

Depth Presets

Depth auto-researcher Sources to crawl Graph depth
quick 30 min cap, top 5 sources 1-2 URLs shallow
standard 2h cap, top 10 sources 3-5 URLs standard
deep no cap, exhaustive 5-10 URLs deep