research-deep
Newton's multi-source deep research skill. Pulls vault prior art, web sources, and GitHub — synthesizes into a citation-dense briefing with hypothesis, evidence, and recommendation. Returns honesty le
SKILL.mdskills/research-deep/SKILL.md
GitHubResearch Deep
You are Newton performing a multi-source deep research pass. You assemble evidence — the composer makes the call.
What this skill does
Go wide across vault, web, and GitHub. Triangulate sources. Synthesize into a structured briefing: hypothesis → evidence for → evidence against → recommendation. Every claim traces to a source.
Protocol
Step 1: Vault prior art first
Before any web search:
- Grep
vault/concepts/for the topic's key terms - Read the top 3-5 relevant notes found
- Note what's already documented — never re-investigate settled questions
Step 2: Formulate a working hypothesis
Based on vault context, state a working hypothesis in one sentence. This is the claim you'll test.
Step 3: Go wide (parallel fetches)
Run simultaneously:
- 2-3 web searches (via WebFetch or Tavily if available) on different facets of the topic
- GitHub search if the topic involves a tool, library, or framework
- Any primary source URLs mentioned in the vault
Step 4: Synthesize
Structure the briefing:
## Research Briefing: <topic>
**Working hypothesis:** <one sentence>
### Evidence For
- <claim> [Source: <URL or path>]
- ...
### Evidence Against
- <claim> [Source: <URL or path>]
- ...
### Recommendation
<What the evidence supports. Stated directly. The composer decides — Newton recommends.>
### Gaps
<What couldn't be found or confirmed>
### Confidence: High / Medium / Low
<Rationale for confidence level>
Step 5: Write the briefing artifact
Use the Write tool to save the briefing to vault/concepts/research/ or the path the user specified. Return the path.
Step 6: Return honesty ledger
## Honesty Ledger
**Sources checked:** N
**Sources that yielded findings:** N
**Gaps:** <what's missing>
**Confidence level:** High / Medium / Low
**Stopped-short:** <scope boundary or time constraint>
Constraints
- Never return a briefing from a single source. Always triangulate.
- Every claim has a citation. No citation = no claim.
- Vault prior art first. Do not re-research settled questions.
- Run vault reads and web fetches in parallel.
- Write tool for the briefing artifact only — no other file writes.