Source: call transcript Michael shared 2026-08-05 plus the JOLT AGENTS BUILD STARTING POINT and JOLT Basic Agent Builds Notion pages (35f8ab5fb6c780a082ddff9e419f4527, 3a08ab5fb6c780e299b1ee8e45433fe2). This doc extracts the three items the Hermes-native JOLT rebuild must explicitly carry over. Everything else from those sources is already operationalized in ~/.jolt/workspace/knowledge/ short-form-best-practices.md and per-client content-dna.md files.
1. Trend detection thresholds (Tatem’s calibrated numbers)
Encode as named constants with attribution in the Trent replacement; these are practitioner-calibrated, not arbitrary:
- Audio trend forming: audio at roughly 12,500 uses AND the top 3 videos under it posted within the past 5 days = trend likely breaks in the next 24-48 hours. Tatem: “if the top three videos under that audio have been posted within like the past five days and the audio is roughly at 12.5 thousand uses… that’s probably going to turn into a trend really soon, like in the next 24 to 48 hours.”
- Profile outlier signal: a video that dramatically outperforms ITS OWN creator’s baseline is the idea signal. Vanity metrics mean nothing on big accounts. Tatem: “it does mean something if that person has a low amount of views on most videos, but that one video popped off.” Notion operationalization: 3x+ the creator’s average views (profile search) or 3x+ the result-set median (hashtag/keyword search).
- Trend window passed: audio at 50k+ uses, or the anchor video 6+ days old. Flag as late rather than surfacing as urgent (Brian Storm rule in the Notion doc).
- Recency dominates: “you have to go with what’s happening like that week.” Default search timeframe 7 days.
2. Voice capture method (speaking voice beats writing voice)
For video scripts, written samples are the WRONG source. Two-part capture per client, make it a standard onboarding step in the Hermes-native flow:
- Informal brain dump: “set a five minute timer and type out your thoughts on a random topic… like you’re talking to your mom or your girlfriend… instead of talking to a business person.” That is the only WRITING sample ever needed.
- 2-3 meeting transcripts (kickoff call required; sales/team calls optional): the primary voice source. Ben in the call: “If we could ask for transcripts of like three different transcripts of their meeting content… that would nail their voice pretty well.” Tatem: “the way that you speak and the way you type is just two different things.”
Downstream test (Paige standard): client reads the script off a phone teleprompter once and it flows; stumbling = wrong register. Script structure: HOOK (1 line) / CREDENTIALS (1-2 lines, reason to stay, not a bio) / BODY / CTA. Default 30s at natural pace = roughly 75-90 words.
3. TikTok scheduling constraint (platform-specific hard rule)
Tatem, via the STARTING POINT page: schedule inside the native apps only; “never use external tool because it hurts performance.” The Hermes-native scheduler for JOLT must therefore be native-scheduling or manual-posting only for TikTok. Do NOT generalize this to Megaphone Studio: X publishing there uses X’s own API as the authorized user, which is not the third-party-scheduler pattern this warning is about.
Also noted (lower priority, from the same sources)
- Video Analyst concept: pre-publish feedback on recorded videos (tone, lighting, pace, wardrobe) benchmarked against what performs in the niche; 250 views cited as TikTok’s minimum distribution floor.
- Captions: humanize hard; “less is more”; emojis can be enough; up to 5 hashtags still matter on TikTok and trending ones are surfaced in-app.
- Approver flags: unbranded product promotion, profanity (per client profile), overly political content (suppression/boost-block risk).
- Down the line: carousels on TikTok + IG as a distinct format lane; IG Trial Reels when a client already has an IG audience.