Fact-checking
81
links verified in a 32-page client deck, built in 3 h 10 min.
AI drafts sound right and can be wrong.
Every claim is checked against its primary source before anything goes to a client. I hold my own analyses to the same standard: in a market analysis a second agent tries to refute every claim, and the refuted theses stay in the report.
How it works
- Claude Code draftsfrom my brief
- List every claim and link · Claudeone row per claim
- Open each source · Claudethe source, never my own deckFrom: primary sources: registers, statistics, official company pages
- One verdict per claim · Claudeconfirmed, adjusted, did not hold
- I read the tablefix or cut what failed
- Regression checkrefuted sentences must not returnTo: the deck goes to the client with its links verified
What broke and what I changed
After a deck went out, third parties refuted several of its claims.
One had passed as “fine” because I had compared it with my own deck instead of the source. Now every claim is checked against the primary source, and every refuted sentence goes into a regression test so it cannot come back.
Checking a new deck exposed mistakes in my reference list and in a deck that had already gone out.
The reference list is fixed, and a regression test covers it.
Guardrails
Nothing goes to a client until every claim has a verdict and a source.
I read the verification table myself before a deck is sent.
A claim is checked against its primary source, never against another document of mine.
References come only from the approved public list.
A regression test fails if a sentence that was once refuted comes back.
Facts are kept apart from estimates.
An estimate is labelled as one, and quotes come only from sources I have saved.
Numbers
- 81
- links verified in one client deck
- 32
- pages in that deck
- 3 h 10 min
- from an empty folder to the finished PDF
- 4 rounds
- in my market analysis, with a second agent trying to refute every claim