← Resources · Research · Guide
Research fails at the handoff, not at the gathering. Forty pages of notes and a team that still doesn't know what to build. This is how I compress a pile of talks, papers and threads into one page someone can act on — and what I do to keep it from being confidently wrong.
The method has three passes: collect wide, verify hard, cut to one page. They're separate on purpose. Verifying while you gather makes you stop at the first plausible source; cutting while you verify makes you keep the claims you already wrote nicely.
Sweep from several angles at once, because each angle is blind to what the others find:
For a talk or interview, work from a timestamped transcript, not from memory or a summary. Every claim you carry forward keeps its timestamp — so a disagreement later takes thirty seconds to settle instead of a re-watch.
One line per claim, in this shape, from the first minute:
CLAIM | Model training costs fell ~40% year over year
SOURCE | [name], [type: filing / talk / post], [date], [link or 14:32]
KIND | fact · forecast · opinion · anecdote
STATUS | unverified
KIND is the field people skip and then regret. A forecast dressed as a fact is the single most common way a brief goes wrong — nobody lied, the tense just quietly changed somewhere between the transcript and the slide.
Now go back and try to break your own list.
Every load-bearing claim needs two sources that don't derive from each other. Three articles citing the same press release is one source wearing three hats — and it's the failure mode that feels the most like thoroughness.
| Prefer | Over |
|---|---|
| The filing, the paper, the primary document | The article about it |
| The full transcript | The clip |
| The dataset | The chart made from it |
| The person saying it themselves | Someone reporting what they said |
On-screen titles and channel descriptions are marketing. Cross-check who someone actually is and what they're actually selling before you weight their claim — a founder's forecast about their own market is data about the founder.
When two good sources disagree, the honest output is "sources disagree, here's the range" — not the midpoint, and not the one that fits your narrative. Every claim ends as one of four: verified · disputed · unverified · false. "Unverified" is a legitimate thing to ship. A fabricated certainty is not.
The point of verification isn't confidence. It's knowing precisely where your confidence ends.
One page is the constraint that does the work. It forces a decision about what actually matters, and that decision is the value you're delivering — not the notes.
Everything else — the transcript, the full claim list, the dead ends — goes in an appendix or a linked doc. It isn't wasted. It's the thing you reach for when someone challenges line 3, and it's why the one page holds up.