Cited Longform Researcher

Long-form writing earns trust one verifiable claim at a time and loses all of it with one fabricated citation. This skill runs research and drafting as a single disciplined pipeline: questions first, sources ranked by proximity to the evidence, every claim logged in a ledger before it may appear in prose, and a verification pass that treats the finished draft as hostile testimony.

When to use this skill

• A whitepaper, briefing, essay, or investigative piece needs writing with checkable citations
• Existing research notes must become a draft without the facts drifting in transit
• A draft full of unsourced claims needs auditing before publication
• A contested topic requires deliberately collected counter-evidence, not just support
• Someone asks for "a well-researched piece on X" and actually means it

Workflow

1. Decompose the topic into five to ten research questions whose answers would constitute the piece. Write them down; they are the acceptance criteria for the research phase.
2. Rank sources by proximity to the evidence: primary documents and datasets first, then peer-reviewed work, then institutional reports, then quality journalism, then everything else. Record each fact's tier — a claim resting entirely on tier four gets stated more carefully.
3. Build the claim ledger as you read, one row per claim: the claim, the source, a locator (page, section, table), quote or paraphrase, confidence, tier. No reading without logging; a fact you cannot relocate is a fact you do not have.
4. Triangulate the load-bearing claims. Anything the argument would collapse without needs two independent sources, or an explicit single-source flag that survives into the published text.
5. Run the counter-evidence sweep before drafting: search specifically for the strongest sources that disagree with the emerging thesis, and log them with the same care. A piece that has never met its opposition is advocacy wearing research's clothes.
6. Draft from the ledger, never from memory. Every statistic, date, name, and quotation is pasted from its ledger row, not retyped — retyping is where numbers mutate.
7. Citation pass: walk the draft claim by claim and match each to a ledger row. Unmatched claims get sourced or cut, without exception.
8. Quote audit: quotation marks only around words that appear verbatim in the source, with ellipses marking every omission. Check each against its locator, character by character.

Output format

• Inline citations in the user's preferred convention — numbered, author-date, or footnotes — applied consistently, plus a complete source list at the end.
• The claim ledger ships alongside the draft as an appendix or companion table:

| Claim | Source | Locator | Quote/Para | Tier | Confidence | Corroborated by |
|-------|--------|---------|------------|------|------------|-----------------|

• Load-bearing single-source claims are marked in the text itself ("reported only by..."), not just in the ledger.

Guardrails

• No citation to anything not actually opened and read at the cited location. Citing an abstract as if it were the paper, or a headline as if it were the study, is fabrication with extra steps.
• Paraphrase may change the words, never the strength: "suggests" does not become "shows" on the way into the draft.
• Statistics travel with their year, population, and method — a number stripped of its context is a different claim wearing the same digits.
• When credible sources conflict, the text says so and shows both; silent averaging is invention.
• If the research does not support the requested thesis, report that finding to the user. The evidence gets a vote, and it votes last.