Three patterns after about forty episodes:
1. It turns jokes into claims. A guest says 'we basically invented email' as a joke and it becomes a timestamped highlight. 2. It over-trusts names it hears once and spells them confidently wrong. 3. It writes chapter titles that summarise rather than invite.
Fixing 1 and 2 with a 'flag, do not assert' rule. Still stuck on 3.
For 3, give it five chapter titles you liked from other shows and say 'match this energy, not this content'. Examples beat adjectives for us.
On names: we feed the guest list and their companies before the transcript. Spelling errors dropped to almost none.
Same three problems with video transcripts. For the jokes we ask it to mark anything said with a laugh in the transcript as 'tone: joke' before summarising.
On chapter titles: give it the question the chapter answers, not the content. 'Why did the launch slip' invites; 'Launch timeline discussion' does not.
We use the show notes agent for community digests too. The 'flag, do not assert' rule is the single most useful line in our brief.
Guest list fixed the names. The example titles helped too, about half are usable now, up from none.