Pj2.dZddlZddlZddlZddlZddlZddlZddlZddl m Z m Z ejjejjeZdZdZdZdZdZd Zd d d d ZdZdZddZddZdZedk(reyy)udBuild one ~20-minute themed episode script from a collect_day.py material file. Segments: cold open -> radar briefing -> day anchor (-> Friday zeitgeist) -> outro. Each segment is written by the chosen writer with its own word budget, then concatenated into one script file + .meta.json in the exact format tts_vibevoice.py already renders. Writers: local — gemma on :8000 (the production default) claude — `claude -p` (claude-sonnet-5), no tools; same prompts, same material Usage: build_episode.py --day thu|fri|sat --writer local|claude --material material.json --script-out script.txt N)datetimetimezonez)http://localhost:8000/v1/chat/completionszgemma4-26b-a4bzclaude-sonnet-5uSPEECH-SAFETY RULES (this text goes directly to a text-to-speech engine): - Plain spoken prose ONLY. No markdown, no asterisks, no headers, no bullet lists. - Never write citation brackets like [1]; name authors naturally or say "one earlier study". - Never write URLs, arXiv IDs, file paths, or version strings. - No mathematical notation or symbols: say "F one score" not "F1", "about eighty-seven percent" not "87%". Expand acronyms on first use. Numbers under twenty as words. - Never attempt to pronounce Lakota or other Indigenous-language words; describe them ("the Lakota word for plums") or spell them letter by letter, naming diacritics. Machine pronunciation of these words teaches learners errors. This is ONE SEGMENT of a continuous episode: no greeting, no sign-off, no "welcome back" — open with at most a one-sentence transition, unless the segment instructions say otherwise. - Never narrate the show's structure or your own process: no meta commentary about the script, the sources, or what you are about to do; no filler signposts like "let's dive in" or "that wraps up this segment". Transition through content, not announcements.zThe listener builds community-sovereign language AI for Lakota (OCR of legacy texts, RAG grounded in community-approved books, per-stage evaluation, community-governed data) and listens while driving.uThis episode has TWO hosts in conversation: - Speaker 1 leads: frames each topic, carries the narrative, addresses Mason directly. - Speaker 2 is the analyst: asks the questions a sharp listener would ask, adds caveats and context, pushes back when a claim is thin. FORMAT: every turn is ONE line starting with exactly "Speaker 1:" or "Speaker 2:" followed by that host's words. Never put a speaker tag mid-line. No names, no stage directions, no other formatting. Make it a real conversation — reactions, short interjections, genuine questions, occasional disagreement — never two alternating monologues. Turns should vary in length; most under sixty words.)u3Roundup — Native language tech & data sovereigntyz0weekly Native research and language-tech roundup)zGeoAI deep-dive + AI zeitgeistz5a geospatial AI deep-dive plus the week in general AI)zWeek in reviewz0synthesis of everything the radar kept this weekthufrisatctdt|dzdzd|dgd}tjj t t j|jddi }tjj|d 5}t j|d d ddjcdddS#1swYyxYw)Ng333333?g@,user)rolecontent)model temperature max_tokensmessagesz Content-Typezapplication/json)dataheadersX)timeoutchoicesrmessager ) LLMinturllibrequestRequestLLM_URLjsondumpsencodeurlopenloadstrip)prompt budget_wordsbodyreqrs podcast/build_episode.py write_localr*?sC s@R>> before it and a line containing exactly <<>> after it. Nothing but spoken prose between the markers.claudez-pz--modelz--disallowedToolszARead,Write,Edit,Bash,Glob,Grep,WebSearch,WebFetch,Agent,TodoWriteTiz/tmp)inputcapture_outputtextrcwdrzclaude -p failed: z <<