research

name: research description: "Use when needing to research a topic from the web — scans X and web for state-of-the-art thinking from top voices, synthesizes into bullet points with citations, and optionally ingests into vault" context: fork

Vault Location

Vault path: Read from ${CLAUDE_PLUGIN_ROOT}/brain-os.config.md

/research [topic] [--depth quick|deep] [--ingest] [--platforms web,x,youtube,reddit] [--voices person1,person2]

Behavior

Research agent optimized for tracking cutting-edge thinking from top voices: plan → person-specific + topic searches across platforms → synthesize into cited weekly digest → optionally ingest to vault.

Default behavior: X-first, voice-driven. Generic web articles are low-signal. Prioritize original thinking from researchers, engineers, and founders who are actually building.

Parameters

Voices Registry

Maintain a living table of key voices in {vault}/knowledge/research/voices.md. Auto-populated on first run, user can add/remove anytime.

Default Voices (auto-seeded for AI Engineering)

| Person | Affiliation | Focus | X Handle |
|--------|------------|-------|----------|
| Andrej Karpathy | ex-OpenAI/Tesla | Autoresearch, agentic engineering | @karpathy |
| Harrison Chase | LangChain | Agent harness patterns, flow engineering | @hwchase17 |
| Swyx (Shawn Wang) | Latent Space | AI Engineer role, agent engineering | @swyx |
| Simon Willison | Independent | LLM tooling, practical AI, MCP | @simonw |
| Jediah Katz | Cursor | Context engineering, dynamic discovery | @jediahkatz |
| Tobi Lütke | Shopify | AI-first org transformation | @tobi |
| Dex Horthy | HumanLayer | 12-Factor Agents, production patterns | @dexhorthy |
| Anthropic team | Anthropic | Claude, MCP, multi-agent research | @AnthropicAI |
| OpenAI team | OpenAI | Codex, deep research, reasoning | @OpenAI |

User adds voices via: /research --add-voice "Name, Affiliation, Focus, @handle"

Phase 1: PLAN

1. Read {vault}/knowledge/research/voices.md — load tracked voices
2. Merge with any --voices override
3. Decompose topic into 2-4 angles (not generic subtopics — think: paradigm shifts, tools, architecture patterns, org transformation)
4. For EACH angle, generate:
   • Voice-specific queries: "{person_name} {topic_angle}" site:x.com
   • Topic queries: "{topic_angle} 2026 latest" site:x.com
   • Web deep-dives: for blog posts, papers, repos linked from X posts

Phase 2: SEARCH

Search Strategy — X-first

Round 1: Voice tracking (highest signal) For each tracked voice relevant to the topic:

WebSearch: "{person_name} {topic}" site:x.com

This catches their latest takes, threads, and announcements.

Round 2: Topic scanning

WebSearch: "{topic_angle} site:x.com 2026"
WebSearch: "{topic_angle} new approach paradigm 2026"

This discovers new voices and conversations not yet in registry.

Round 3: Deep-dives (web) For high-signal X posts that reference blogs/repos/papers:

WebFetch: extract key points from linked blog/repo
WebSearch: "{specific_concept_from_X} explained 2026"

Round 4 (if --platforms includes youtube/reddit):

WebSearch: "{topic} site:youtube.com podcast 2026"
WebSearch: "{topic} site:reddit.com discussion 2026"

Round 4.5: video / podcast URLs → /transcribe-video FIRST

If the topic search surfaces a YouTube watch link, podcast episode, or audio URL with direct verbatim signal (e.g. a key talk by a tracked voice), DO NOT extract quotes from same-cycle aggregator articles. Aggregators conflate quotes from different talks by the same person — citing them as primary source is a recurring failure mode.

Instead, invoke /transcribe-video (skill) before quote extraction:

bun ${CLAUDE_PLUGIN_ROOT}/scripts/transcribe-video.ts <URL> --out {vault}/knowledge/research/findings/{slug}/

This writes _transcript-verbatim.md to the findings folder. ALL subsequent finding files cite quotes that grep into this file; aggregator articles inform structure (which topics matter) but never source quotes. Source-tag findings as [primary — verbatim] not [paraphrase].

When the topic is technical and aggregator articles agree on framing, this step is optional. But when the user pushes back ("did they really say that?") or quotes are load-bearing for content extraction, transcribe-video is mandatory.

See [[skills/transcribe-video/SKILL.md]] for usage contract, two transcription paths (auto-subs vs whisper-cpp), and known-typo replacement table.

Quick mode

5-8 WebSearch calls: focus on voice tracking + topic scanning.

Deep mode

Spawn up to 3 subagents in parallel:

• Agent 1: Voice tracking (all registered voices)
• Agent 2: Topic angle scanning on X
• Agent 3: Deep-dive web/youtube/reddit Each writes to {vault}/knowledge/research/reports/research-findings-{slug}-{n}.md

Phase 3: SYNTHESIZE

1. Cluster findings by theme (not by person)
2. For each finding: one bullet, one line, one source link
3. For breakthrough ideas: include a short quote (< 15 words) from the original post
4. Build a Voices table with each person's current focus
5. End with Open Questions

Output Format

# {Topic}: State of the Art — Week of YYYY-MM-DD

## Paradigm Shifts
- **{shift name}** — {one-line explanation} ([Person on X](url))
- ...

## Key Approaches & Patterns
- **{pattern name}** — {what it is, why it matters} ([Source](url))
- ...

## Tools & Infrastructure
- **{tool}** — {what's new} ([Source](url))
- ...

## Org Transformation Signals
- **{company/person}** — {what they did/said} ([Source](url))
- ...

## Key Voices & Current Focus
| Person | Affiliation | Current Focus | Follow |
|--------|------------|---------------|--------|
| ... | ... | ... | [@handle](url) |

*Add your own voices to this table as you discover them.*

## Open Questions
- {what remains unclear, contradictory, or worth watching}

---
*Sources: {n} searches across {platforms} | Voices tracked: {n}*

Save to: {vault}/knowledge/research/reports/YYYY-MM-DD-research-{slug}.md

Phase 4: INGEST (optional, only with --ingest)

If --ingest flag is set:

1. Convert key findings into atomic vault notes
2. Save to {vault}/knowledge/research/findings/{slug}/
3. Each note = one finding, tagged source: "research-{date}"
4. Link to related vault concepts if they exist

Research Note Template

---
source: "research-{date}"
topic: "{topic}"
url: "{source_url}"
voice: "{person_name}"
signal: {high|medium|low}
tags: [{tags}]
---

# {Finding Title}

{1-2 sentences: the finding, precise and factual}

## Implication
{Why this matters — one sentence}

## Related
- [[related-vault-note-if-exists]]

Integration with Other Skills

Weekly Usage Pattern

# Monday morning: scan what happened last week
/research "AI engineering" --depth deep --platforms x,web --ingest

# Midweek: deep dive on something specific from Monday's scan
/research "autoresearch autonomous loops" --depth deep --platforms x,web,youtube

# Ad-hoc: check a specific person's latest thinking
/research "context engineering" --voices "Jediah Katz, Anthropic"

Examples

/research "AI agents state of the art"
/research "agentic engineering" --depth deep --platforms x,web,youtube
/research "MCP protocol ecosystem" --voices "Simon Willison, Anthropic" --ingest
/research "AI transformation enterprise" --platforms all --lang vi
/research --add-voice "Yann LeCun, Meta, AI skeptic counterpoint, @ylecun"

Guardrails

• Never fabricate sources — every bullet must have a real URL from WebSearch results
• X-first — original posts from builders > blog rewrites > generic articles
• Recency bias — prefer this month's sources; flag anything older than 3 months
• Signal hierarchy — primary source (X post, blog, repo) > analysis > opinion > news summary
• One line per bullet — if it needs a paragraph, it's not distilled enough
• Quotes < 15 words — only for truly original phrasing worth preserving
• Copyright — never reproduce long quotes, only short factual extractions
• Max searches: quick=8, deep=20

Outcome log

Follow skill-spec.md § 11. Append one line to {vault}/daily/skill-outcomes/research.log:

{date} | research | {action} | ~/work/brain-os-plugin | {vault}/knowledge/research/reports/{date}-research-{slug}.md | commit:{hash} | {result}

• action: quick, deep, or add-voice
• result: pass if report written with citations, partial if low signal, few sources, or some platforms failed, fail if search errors
• Optional: args="{topic}", score={sources_found}
• Conflict recovery: if git pull --rebase fails, run git rebase --abort. Log partial if the report file was written to disk, fail if no report was produced. Use commit:conflict in the log line.