---
name: "research_methods"
description: "Conduct rigorous research: distinguish primary vs secondary vs tertiary sources, evaluate evidence quality with a structured rubric, chain citations to foundations, and cite correctly. Use when answering factual questions, writing research briefs, drafting papers, or verifying claims."
metadata: { "includeInPrompt": true }
---

# Research Methods

## Purpose
Give an agent the working habits of a competent researcher: find evidence, weigh it, cite it, and be honest about what remains unverified. Use whenever a task needs facts it can stand behind — research briefs, literature reviews, due-diligence, technical investigations, academic writing.

## Workflow

### 1. Frame the question
Write one sentence: the exact claim or question to answer. Split multi-part questions into separate claims — each gets its own evidence, not one blended answer.

### 2. Collect sources (lateral reading first)
- Prefer **primary sources**: the original study, dataset, filing, statute, official docs, raw measurements, first-person account. Secondary sources (news reporting on a study, biographies, reviews) are one step removed; tertiary (encyclopedias, Wikipedia, handbooks) are maps to sources, never citations.
- **Lateral reading**: when weighting source X, search the publisher/author independently — don't trust their self-description.
- Never treat ten re-summaries of one press release as ten sources. Diversity of source *type* beats diversity of source *count*.
- For any claim that drives a decision, get **two genuinely independent sources** (syndicated copies and sources relying on the same material do not count). One authoritative source suffices for background.

### 3. Evaluate each source
Run the CRAAP rubric (librarians at CSU Chico, 2004) — score each criterion 0–3:
**C**urrency · **R**elevance · **A**uthority · **A**ccuracy · **P**urpose.
Total ≥ 11: high confidence · 7–10: moderate · ≤ 6: skepticism. See `references/source-rubric.md` for the full worksheet, red flags, and the source-quality ladder.

### 4. Chain citations
Trace references 2–3 levels deep: L1 = directly relevant sources, L2 = the foundational work they build on, L3 = intellectual context. If every L2 source cites the same L3 paper, that paper is probably foundational — read it. When a secondary source makes the key claim, follow its citations back to the primary when possible.

### 5. Keep a source trail
Log every source that carries weight: id, type (primary/secondary/tertiary), author or publisher, date, URL/locator, claims it supports, its limitations. Template in `references/source-ledger-template.md`. A source mentioned only in a trailing list without being referenced in the body is not a citation.

### 6. Cite correctly
- **Cite when**: quoting, paraphrasing or summarizing, stating anything highly debatable, using data/images/procedures/opinions not your own.
- **Common knowledge** (what an educated person would know, could observe, or readily find in a textbook/encyclopedia) usually needs no citation — but it's context-dependent. When in doubt, cite.
- If you cannot find the source your data came from, do not use the data.
- Match the requested style (APA, MLA, Chicago) or default to: `Author. (Year). Title. Publisher. URL`.
- Distinguish source statement from your inference. Label inference explicitly: `Inference: …`

### 7. Write the answer
Lead with the answer, then the evidence. If credible sources conflict, report the conflict and explain the difference — never silently pick one. Close with a `## Sources` section: one fetched source per line, `N. [Title](URL)`. Every listed URL must have been actually retrieved this turn.

## Output Contract
- The question answered in one sentence, up front.
- Findings with per-claim attribution (source + date where time-sensitive).
- Explicit flags for conflicts, inference vs. statement, and evidence gaps.
- `## Sources` list: only sources retrieved this turn, one per line.
- The source trail/ledger preserved in the working directory.

## Operating Rules
1. **Every factual claim must be verified via web search or a primary source.** Never invent numbers, dates, quotes, or events. Unverifiable = cut it or frame as opinion. No confident unverified claims, ever.
2. When retrieval fails or a fact cannot be verified, say so. A stated evidence gap is better than a confident guess.
3. Weight advocacy sources by their stake in the answer; a trade association's report on its own industry's product needs corroboration from a non-aligned source.
4. Prefer primary sources over aggregators, SEO content, or AI-generated summaries — those are leads only.
5. Answer the actual question asked. Don't turn a focused research request into a broad survey unless the extra material directly helps answer it.
6. For time-sensitive facts (prices, versions, laws, schedules), record the publication/access date with the claim.
7. Never self-plagiarize across papers and never submit one piece of work for two purposes without disclosure.

## References
- `references/source-rubric.md` — CRAAP worksheet, red flags, source-quality ladder.
- `references/source-ledger-template.md` — copy-paste ledger template + worked example.
- `references/citation-checklist.md` — what-to-cite / what-not-to-cite rules per style-guide conventions.