Structured Web Research

Open-ended web research fails by drift: twenty tabs, three contradictory numbers, and a summary that quietly sides with whichever source was read last. This skill runs research as a structured pipeline — decompose the question, rank source types before reading anything, extract claims into a table as you go, triangulate whatever is load-bearing, and synthesize only from the table — so the final answer carries its evidence with it.

When to use this skill

Market, competitor, technical, or policy questions that a single lookup cannot settle
Claims that will be repeated to others — in a report, a decision memo, a public post
Reconciling sources that disagree about a number or a fact
Building a fact base that another agent or analyst will reuse later

Workflow

Decompose the question into 3-8 sub-questions, each answerable by a source in principle. Mark each one fact (has a true answer), estimate (a range is expected), or judgment (the answer is an argument). The type sets how much triangulation it deserves.
Rank source classes before searching, so quality is a decision rather than an accident: primary documents (filings, specifications, standards, court records, raw data) outrank official statements by the party concerned, which outrank reporting with named sources, which outranks expert commentary, which outranks aggregators and forums. Note the classes you expect to need per sub-question.
Search in passes and log as you go: broad orientation first, then targeted queries per sub-question, then adversarial queries built to find disconfirmation — "<claim> criticism", "<number> overstated", "<approach> problems". A log of queries tried, including dead ends, prevents re-treading and exposes blind spots.
Extract into the claims table at read time, never from memory afterwards. One row per claim: the claim, the source, its class and publication date, what the source's own evidence actually is (their data? a citation? bare assertion?), and a confidence.
Chase citations to their origin. When three articles cite the same survey, that is one source wearing three hats. Record the origin as the source; count the echoes as reach, not confirmation.
Triangulate every load-bearing claim with at least two genuinely independent sources — independence means different origin, not different URL. Where sources conflict, keep both rows and write the best explanation for the gap (different definitions, dates, incentives) instead of averaging it away.
Date-stamp everything and match freshness to volatility: prices and headcounts rot in months; physical constants do not. Flag any claim whose source predates a known relevant change.
Synthesize from the table only. Every sentence in the write-up traces to rows; whatever cannot is labeled as inference and argued, not asserted. State confidence per conclusion and name what evidence would change it.

Output format

## Research: <question> — <date>

### Answer
<direct answer in 3-6 sentences, with confidence stated>

### Findings by sub-question
<sub-question>: <finding> [S1, S3] — confidence: high | medium | low

### Claims table
| Claim | Source (class, date) | Their evidence | Independent? | Confidence |

### Conflicts and unknowns
<disagreements kept visible, plus what would resolve them>

### Sources
[S1] <title, publisher, date, link>

Guardrails

A company's statements about itself are claims by an interested party — usable, never sufficient alone for a load-bearing fact
Fluency is not sourcing: a confident summary with no rows behind it is the exact failure this skill exists to prevent
Absence of evidence is reported as absence — "no public figure found" — never filled with a plausible guess
Quotes and numbers are captured at read time; links rot and pages change under you
Time-box the orientation pass; depth goes to the load-bearing sub-questions, not the interesting ones

Signs the research is done

Every load-bearing claim has two independent rows, or carries an explicit single-source caveat
New sources are repeating existing rows rather than adding facts
The conflicts section explains disagreements instead of hiding them
You can name what you would look at next — and say why it can wait