Skip to content

Investigations

github-actions[bot] edited this page Nov 21, 2026 · 4 revisions

receipts starts from a story and finds his old clips; clippa investigate starts from the person and looks for the story. It reads an archive subject's own transcripts and comes back with leads: claims that changed, numbers that moved, teases never paid off, topics that suddenly took over. Leads are research for you, never uploaded or posted.

clippa investigate facts jer --since 2026-09-01 --dry-run   # items, hours, estimated cost
clippa investigate facts jer --since 2026-09-01             # Haiku reads each item for checkable facts (~$0.20 for 43 h)
clippa investigate leads jer [--about "lawsuit"] [--confirm] # threads found in code, Sonnet writes the leads (~$0.10)
clippa investigate check jer [--lead L03] [--trail]          # is anyone else on it? the search swarm, per lead
clippa investigate list jer                                  # kind, confidence, coverage, receipts
clippa investigate show jer L03                              # the story and the timeline, quote and link per receipt
clippa investigate confirm jer [--lead L03]                  # test leads against more of the archive (`archive ask`)
clippa investigate run jer                                   # facts --new -> leads -> check, total cost
clippa investigate post jer [--new-thread]                   # the kept leads to Discord #investigations

facts keeps only facts whose quote is found in the item's words (in the part it was read from, with its numbers and any negation unchanged), and the receipt keeps his wording, so every receipt in a lead is something he said, linked to the moment. Each item's facts are cached (archive/jer/investigate/facts/<item>.json) until the prompt version or the transcript changes, so a rerun reads only new uploads. The whole Jer catalog (2,731 h) is about $14 once; quote the --dry-run figure before running it. It counts the prompt, the schema and the header that every call repeats (about 1,800 tokens), one call per [investigate] chunk_s of video and at least one an item, so a catalog of short clips costs much more per hour than long videos. leads builds facts.sqlite and finds candidate threads with no model call: a trend, a contradiction (affirmed, then denied), a shift (a number that changes), a dropped thread (checked against the full transcripts) and a convergence (personal remarks from several videos close together); Sonnet reads each one and writes the lead or rejects it, and the rejected ones and why are kept in leads.json too. check runs the search swarm on each kept lead, seeded with the lead's search words on YouTube (titles only), the web and Reddit (Brave, the clippa cast key; without one, Reddit's own feed, which often refuses) and our own streams' segments; its planner adds searches and its workers follow up, within [investigate] swarm (small: ~$0.01-0.03 a lead). With his own video titles, Haiku labels each lead fresh, partly covered or covered. The results and verdict are in investigate/swarm/<lead>.json, the task tree in <lead>.trail.json (--trail prints both). Edit leads.json to flip keep, change a headline or add notes; leads.md is the readable dossier. Everyone's names in clippa people merge their spellings, and the names of the person whose archive it is count as him. Put other spellings Whisper gets wrong, or other names for him, in investigate/aliases.json ({"Jeremy": "self", "Hassan Piker": "Hasan Piker"}), which wins over clippa people. A run stops before it would spend more than [investigate] budget_usd. Whisper can mishear a number, so a "shift" is worth a listen at the link before you use it.

With [investigate] discord = true, leads, check and confirm end by posting to [investigate] discord_channel (#investigations, made if missing; a forum channel of that name works too): one thread per investigation, that is per leads run, named "jer · 2026-11-03" (with the --about focus when there is one), and one short message per kept lead: headline, kind, confidence, coverage, how many receipts, the story's first sentence and a link to the lead in the war room. A later check or confirm edits those messages rather than posting again; a new leads run starts a new thread. The ids are in investigate/discord.json. A post that fails says so and leaves leads.json as it is; clippa investigate post jer retries. The war room's Investigations window shows each subject's dossier (leads.md, rendered), newest first; a Discord link (#inv=jer&run=<made_at digits>&lead=L03) opens it at that lead of that run (an earlier run is read from its leads.<time>.json backup). A lead that stops being kept is marked withdrawn in Discord; a thread or message is only remade when Discord says it is gone.

Setup

clippa investigate needs nothing new beyond a transcribed archive and ANTHROPIC_API_KEY; clippa archive index <subject> (free) lets it check dropped threads against the full transcripts and is needed for --confirm. Its coverage check searches the web and Reddit through the Brave key clippa cast uses (see the cast library above); without one it tries Reddit's own feed, which often turns unsigned clients away. To get the leads in Discord, set [investigate] discord = true: the review bot's token and [discord] guild_id are enough, and the bot makes #investigations if it has Manage Channels (otherwise make the channel by hand; clippa doctor checks the settings).

Clone this wiki locally