research-ops

# Research-ops — the deep-research operating procedure You are the **method, not the topic**. Someone hands you "go find out about X" and you hand back a memo where every load-bearing claim traces to a source, a date, and a confidence tier — a document that survives someone reading it adversarially. The topic changes every time; the procedure does not. Two hard rules, stated up front because everything else hangs off them: - **One search pass is not research.** A single query plus reasoning over the snippets is a guess with footnotes. Real research is an iterative loop — a deep-research run typically reads 20–100+ sources, re-querying as gaps appear, until coverage holds or a budget caps it. - **An unsourced claim is not a finding.** If you can't attach a source and a date, it's an assumption — label it as one or cut it. Human review catches AI errors in roughly 15–20% of research reports, so structure the output so a reviewer can check each claim, never so they have to trust it. ## The loop Research is an ordered loop, not a lookup. Run it in this order; each step has a reason. 1. **Scope** — pin the question down before spending a single search (see *Scope first*). Why: a fuzzy question burns the budget on the wrong sources. 2. **Plan queries** — write 3–6 distinct query angles, not one phrasing repeated. Cover the claim, the counter-claim, and the primary source. Why: you can't triangulate what you only searched one way. 3. **Fan out** — run the searches in parallel; collect candidate sources. Why: breadth first exposes disagreement you'd miss going one source deep. 4. **Fetch & read** — open the actual pages, not the result snippets. Read the primary source, not the blog summarizing it. Why: snippets drop caveats, dates, and numbers. 5. **Re-query on gaps** — every read surfaces a new unknown or a contradiction; feed it back as a new query. Why: this is the part that makes it a loop instead of a list. 6. **Triangulate** — confirm each load-bearing claim across ≥2 independent sources; record where they disagree. Why: triangulation is the foundation of a credible finding. 7. **Synthesize** — write the memo answer-first, every claim carrying `[source, date, confidence]`, with an explicit "couldn't verify" section. Why: the memo is the deliverable; the searches were just inputs. > **Stop rule:** stop when new searches stop changing the answer (coverage plateaus) OR > the budget cap is hit — whichever comes first. Looping forever is not rigor. ## Scope first Refuse to start on an underspecified question. Researching a fuzzy ask produces a fuzzy memo and wastes the search budget. Before the first query, get the 2–3 answers that change which sources are even relevant. ```text Bad (will waste the budget): "What car should I buy?" Good (scopeable): "Best used EV under €25k for a 40km daily commute, bought in Spain in 2026, prioritizing range over trim." ``` The clarifiers that almost always matter: **constraint** (budget / scale / tolerance), **context** (where, for whom, what stack), and **time** ("as of when" — 2026 answers differ from 2023 ones). If the asker can't answer them, ask; don't guess and research the wrong thing. ## Source credibility Not every source counts the same. Restrict to **primary / authoritative** sources where you can — official docs, standards bodies, regulator pages, company release notes, filings, the actual paper. When a claim lives only in secondary commentary, it drops a confidence tier. Pick the credibility check by source type: | Source type | Check to apply | Default confidence | |---|---|---| | Official docs, standards, regulator, filing, release notes | SIFT — Trace to original; you are already at it | High | | Peer-reviewed / scholarly / formal document | CRAAP (Currency, Relevance, Authority, Accuracy, Purpose) | High once it passes | | Trade press / reputable news, corroborated by another | SIFT — Find better coverage, confirm elsewhere | Medium | | Single blog, vendor marketing, forum post, uncorroborated | SIFT — Investigate the source; treat as a lead, not a fact | Low | | AI summary / search snippet | Not a source — open the page it cites | None until traced | **Lateral reading is the non-negotiable move.** To judge a page, leave it: open a new tab and check what others say about the author/org rather than trusting the page's account of itself. A site's "About" page is not evidence the site is authoritative. Full SIFT/CRAAP walkthrough, lateral-reading recipe, and worked tier examples live in [references/credibility-rubric.md](references/credibility-rubric.md). ## Date everything Every non-obvious claim carries a date, because "true" has a shelf life — a 2024 pricing fact or API behaviour may be wrong in 2026. Record **both** dates when they differ: - **Publication date** — when the source was written/last updated. - **Access date** — when you read it (matters for living pages with no clear pub date). Provenance line grammar (the contract the verify gate checks): ```text CLAIM — [Source title](https://url), pub 2026-04-12 / accessed 2026-06-02 · confidence: high ``` If a source carries no discernible date, that's a finding in itself: record `pub: n/a`, keep the access date, and drop it a tier. **Stale → re-verify**: when refreshing an old memo, re-run the loop on the dated claims; don't just copy yesterday's citation forward. Full skeleton and grammar: [references/memo-template.md](references/memo-template.md). ## Verify load-bearing claims A claim is **load-bearing** if the answer changes when it's wrong. Triangulate every one against **≥2 independent sources** — independent meaning they don't both trace back to the same origin (three blogs quoting one press release is one source, not three). When sources disagree, **surface the disagreement; never average it away.** "Source A says X, source B says Y, here's why and which I weight higher" is a finding. Splitting the difference into a number neither source supports is a fabrication. If you genuinely find no disagreement, say so explicitly — silence reads as "didn't check." This is why the memo is structured claim → source → date → confidence rather than as flowing prose: roughly 15–20% of AI research reports contain an error a human catches on review, and they can only catch it if each claim is individually checkable. ## Synthesize The memo is the deliverable. Structure it so the answer arrives first and the evidence backs it up — not a link dump the reader has to assemble themselves. ```markdown ## Answer <the direct answer in 2–4 sentences; the bottom line up front> ## Findings - <claim> — [Source](url), pub YYYY-MM-DD / accessed YYYY-MM-DD · confidence: high|med|low - <claim, triangulated> — corroborated by [A](url, date) and [B](url, date) · confidence: high ## Disagreements - <where sources conflicted, and how you weighted them> — or "none found across N sources" ## Open questions / could not verify - <what you could not source; what's still an assumption; what would settle it> ``` The "Open questions / could not verify" section is mandatory and is not a sign of failure — it's the honest boundary of what the evidence supports. A memo with no open questions on a hard topic is usually a memo that stopped looking. ## Budget & stop rules Cap effort so the loop terminates. Sensible defaults for an on-demand run: - **~5–8 searches** to start; escalate past that only when a real gap or contradiction demands it, not reflexively. - **Fetch what you'll cite**, not everything you find — reading 8 sources well beats skimming 40. - **Stop** when two more searches don't move the answer (coverage plateau) or the cap is hit. Then write — note in *Open questions* anything the budget left unresolved. ## Anti-patterns | Anti-pattern | Why it fails | Do instead | |---|---|---| | One search, then reason over the snippets | That's a guess with footnotes, not research | Run the loop: re-query on every gap | | Citing the search-result snippet | Snippets drop caveats, dates, numbers | Fetch and cite the actual page | | Single-source claim presented as fact | One source can be wrong or biased | Triangulate load-bearing claims across ≥2 independent sources | | Averaging two contradicting sources | Invents a number neither source supports | Surface the disagreement and weight it | | Trusting a page's self-description | "About us" is not evidence of authority | Lateral-read: check the org elsewhere | | Undated citation | A 2023 fact may be false in 2026 | Record pub + accessed date on every claim | | No confidence tier | Reader can't tell a filing from a forum post | Tag high/med/low per finding | | No "could not verify" section | Hides the boundary of the evidence | Always include open questions | | Synthesizing before reading | Conclusion drives the search, not the evidence | Read first, conclude after | | Infinite search, no stop rule | Burns budget, never ships the memo | Cap searches; stop at coverage plateau | | Topic-creep into market sizing | That's a different skill's job | Route TAM/SAM/SOM to ../market-research | | Re-citing a stale memo unchanged | Yesterday's source may be outdated | On refresh, re-verify the dated claims | ## Verify The memo is a checkable artifact, so there's a gate for it. Run it against the produced memo (read-only; it never edits): ```bash ./scripts/verify.sh --path memo.md # check one memo ./scripts/verify.sh --path research/ # scan a directory of memos ``` It asserts the memo has an answer/summary section, that every finding line carries a citation token, that every citation carries a date, that a confidence tier appears, and that an "Open questions / unverified" section exists. A missing or empty target is a SKIP, not a failure. The gate proves the memo is *sourced and dated* — it does not judge whether the answer is correct; that's the capability eval's and your job. ## See also - [../market-research/SKILL.md](../market-research/SKILL.md) — when the question is "how big is the market / who's in it" (TAM/SAM/SOM), not "what's true about X". - [../competitor-watch/SKILL.md](../competitor-watch/SKILL.md) — when you need a standing cadence watch on named rivals, not a one-shot investigation. - [../data-scraper/SKILL.md](../data-scraper/SKILL.md) — when the job is bulk-extracting data from many pages; research-ops *uses* fetched pages, it doesn't own scrape infra. - [../knowledge-ops/SKILL.md](../knowledge-ops/SKILL.md) — to file what you already know into a durable base; research-ops *produces* findings, knowledge-ops *files* them.