thepexcel/deep-research

Deep Research

Better than plain websearch. Faster than full research pipelines.

What makes it better:

1. Landscape Scan — discovers what exists before searching specifics (avoids blind spots)
2. Recency Pulse — catches releases from last 7-30 days, searches upstream providers
3. Parallel search — 2-3 queries at once, multiple angles simultaneously

For rigorous research (hypotheses, COMPASS audit, Red Team, full report) → /deep-research-pro

---

Step 1: CLASSIFY

| Type | When | Do | |------|------|----| | A | Single fact | Search → answer directly | | B | Multi-fact / comparison | SCAN → RECENCY → SEARCH → Synthesize | | C | Judgment / recommendation | B + flag uncertainty + note limitations |

Tiers: Quick (5-10 sources) · Standard (10-20 sources)

---

Step 2: LANDSCAPE SCAN (skip for Type A)

Map what exists before searching specifics. Never use known names in scan queries — you'll miss things that exist but you don't know about yet.

❌ "DeepSeek Qwen performance 2026"   ← only finds what you already know
✅ "China open source LLM list 2026"  ← discovers the full landscape

Queries (parallel):

WebSearch: "[topic] landscape overview [current year]"
WebSearch: "top [topic] list [current year]"
WebSearch: "[topic] all options [current year]"

Extract entity names → split into Discovered (new) vs Confirmed (updated).

---

Step 3: RECENCY PULSE (mandatory for tech/AI topics)

Yearly searches miss releases from last week. Downstream product news lags upstream by weeks.

Map supply chain first: Who makes the underlying tech? → Search them directly.

WebSearch: "[topic] latest news [current month] [current year]"
WebSearch: "[upstream provider] latest release [current month] [current year]"

Example — researching "Microsoft Copilot": Upstream = OpenAI + Anthropic → search both directly, don't rely on Microsoft announcements alone.

Flag anything from last 7-30 days as RECENT or BREAKING.

---

Step 4: SEARCH

Run queries in parallel (single message, multiple tool calls):

WebSearch: "[topic] [current year]"
WebSearch: "[topic] limitations problems"
WebSearch: "[topic] vs alternatives comparison"

Stop when: 3 consecutive searches add <10% new info (saturation) or sources converge on same answer.

URL fallback (403/blocked):

curl -s --max-time 60 "https://r.jina.ai/https://example.com"

Claim confidence:

• C1 (key claims) — need 2+ sources + confidence note: HIGH / MEDIUM / LOW
• C2 (supporting) — citation required
• C3 (common knowledge) — cite only if contested

Never state C1 without citing [N]. If no source found → say so.

---

Step 5: SYNTHESIZE + SAVE

For each key finding, answer:

• แล้วยังไง (So what)? — why does this matter?
• ต้องทำอะไร (Now what)? — action to take?

If sources conflict: flag explicitly — "Source A says X, Source B says Y — likely because [reason]."

Always save output:

research/[topic-slug]-[YYYY-MM-DD].md

Optional: --ingest (flow findings into Obsidian wiki)

If user says --ingest, "ingest to obsidian", "เก็บเข้า vault", "save to wiki", or "update obsidian" — also flow the research into the Obsidian LLM-Wiki at D:\sira-obsidian-sync\ (Karpathy pattern, see its CLAUDE.md):

1. Save the raw synthesis to Sources/articles/YYYY-MM-DD_<topic-slug>.md with frontmatter type: source, status: immutable, source_url list, ingested date. Do not edit this file after saving.
2. Identify wiki entities the research touches — Excel functions, AI models, ComfyUI nodes, business concepts, people, etc.
3. For each entity: either update the existing vault page (D:\sira-obsidian-sync\<entity>.md) with distilled new knowledge + a link back to [[Sources/articles/YYYY-MM-DD_<topic-slug>]], or create a new page if none exists.
4. Update D:\sira-obsidian-sync\index.md with any new page.
5. Update the relevant Hub (MOC) with a link to the new/updated page.
6. Append to D:\sira-obsidian-sync\log.md with format:

   ## [YYYY-MM-DD HH:MM] ingest | <topic>
   Research: [[Sources/articles/YYYY-MM-DD_<topic-slug>]]
   Updated: [[page-1]], [[page-2]], ...
   

7. Flag contradictions in both the updated page and the log entry if new research contradicts old wiki content.

Do NOT ingest automatically — require explicit opt-in per research run. Ingest is compounding, not ephemeral — make sure the research is worth keeping before flowing into wiki.

Delegate to /obsidian-manager if available — it has the full Karpathy ingest procedure and knows the vault schema (obsidian-manager = authoritative executor, deep-research --ingest = trigger).

---

When to Ask vs Just Do

| Ask | Just do | |-----|---------| | Topic too broad → "อยากเน้นมุมไหนคะ?" | Choose search queries | | Interesting sub-topic found | Format output | | Sources conflict on key point | Type A questions |

---

Related Skills

• /deep-research-pro — Full pipeline: hypotheses, QUEST queries, COMPASS audit, Red Team, formal report
• /boost-intel — Stress-test a research finding before making a decision
• /generate-creative-ideas — Cross-industry creative research (no web search needed)