Admin / Persona canon
El Racheteo persona canon
Single source of truth for how the show talks — hosts, the 5-element cold-open, soul-beat density, sanitizer rewrites, per-segment length targets, and per-voice SSML templates. Source file: persona/persona.json (v1).
Hosts
Voice casting. Editing a host's Polly voice or engine flips the TTS Lambda's SSML path automatically — see lambdas/racheteo-tts/handler.ts.
Cold-open canon (5 elements)
Total window: 20-30 s. Hard-enforced by lambdas/shared/prep-validate-blocking.ts:gateIntroBanter and by the intro-mix Lambda's duration gate.
Soul-beat density (the 'sounds human' floor)
Interjection menu: 26 entries. Min 3/90s of narration. Max 2 consecutive dry segments.
"mi gente""mira""ojo""agárrate""agarrate""escúchame""escuchame""te digo""te lo digo""fíjate""fijate""pero mira""óyeme""oyeme""óyeme bien""oyeme bien""y tú""y tu""oye""anda""y dale""dale""ay dios mío""ay dios mio""tú no me creerás""tu no me creeras"
Sanitizer (post-LLM rewrites)
Applied unconditionally to every narration segment, regardless of who authored the text (gpt-4o-mini, Opus, human). Edit here to add a new Polly-generative-mispronounced loanword.
/\breality\s+show\b/gi→reality— 'reality show' → 'reality' (Polly mispronounces the cognate)
/\brealitys?\s+shows?\b/gi→realitys— plural variants
/\b(el|un|ese|este|del)\s+show\b/gi→$1 programa— 'el/un/ese/este/del show' → 'el/un/ese/este/del programa'
/\blos\s+shows?\b/gi→los programas— 'los shows' / 'los show' → 'los programas'
/\bel\s+paquete\b/gi→todo el lío— Phase R: corporate cliché → street equivalent
/\b(vamos a|voy a|vamos)\s+present(ar(les)?|aci[oó]n)\b/gi→te traigo— Phase R: 'vamos a presentar(les)' / 'presentación' → 'te traigo'
/\bla\s+presentaci[oó]n\b/gi→la carpeta— Phase R: 'la presentación' → 'la carpeta' (show vocab)
/\b(en\s+)?resumen\b/gi→al grano— Phase R: 'resumen' / 'en resumen' → 'al grano'
/\bcabe\s+destacar\b/gi→ojo con esto— Phase R: news-anchor cliché → street cue
/\bdestacad(a|o|os|as)\b/gi→tremend$1— Phase R: 'destacada/o' → 'tremenda/o'
/\bdestaca(n|)\b/gi→rompen— Phase R: verb 'destaca/destacan' → 'rompen' (street verb)
/\bla\s+informaci[oó]n\b/gi→el chisme— Phase R: 'la información' → 'el chisme' (show vocab)
/\bun\s+cameo\s+en\s+el\b/gi→un mueble dentro del— Phase S: 'un cameo en el reality' → 'un mueble dentro del reality' (a+el contraction)
/\bun\s+cameo\s+en\s+la\b/gi→un mueble dentro de la— Phase S: 'un cameo en la casa' → 'un mueble dentro de la casa'
/\bcameos?\b/gi→mueble— Phase S: 'cameo' (as person descriptor) → 'mueble'. See glossary.mueble. Hollywood-cameo sense never appears in El Racheteo.
/[,;]\s*[¡!]?\s*aguanta\s+eso\s*[!.,…]?/gi→.— Phase S: ', ¡aguanta eso!' (mid-sentence) → '.' so the orphan comma collapses cleanly into a period
/[¡!]?\s*aguanta\s+eso\s*[!.,…]?/gi→ — Phase S: '¡aguanta eso!' (standalone) → drop. Not Dominican/Boricua street register. Use 'ojo con eso' / 'fíjate' / 'óyeme bien' if a hype beat is needed.
/(^|[\.!\?]\s+)[Aa]y,?\s+(mi|qué|que)\b/g→$1$2— ep22 v4 (2026-05-02): broadened from 'Ay … mi gente' to 'Ay … (mi|qué|que)'. v3 only caught 'Ay, mi gente'; ep22 v3 segment 05 still leaked 'Ay, mi madre,' — same AI-tell pattern, different noun. The class is 'Ay' as a sentence-initial filler-interjection before any of mi/qué/que (covers 'Ay, mi madre', 'Ay, qué cosa', 'Ay, que rico'). Anchored to start-of-text or sentence boundary so we don't strip mid-sentence 'ay' interjections that are legitimate.
/deja\s+que\s+el\s+misterio\s+respire\b/gi→deja que el misterio respire un poquito— ep22 v3 (2026-05-02): Polly Generative reads bare 'respire' (Spanish 3sg subjunctive) as the English /rɪˈspaɪər/ — the user heard an English word at this exact phrase in seg 09. Adding the 'un poquito' tail forces a Spanish prosodic context around 'respire'; combined with the brain/pronunciation.json 'respire' entry, Polly stays in Spanish. Targeted to this collocation so we don't break other 'respire' usages.
/\b(la|esa|esta|toda\s+la)\s+cobertura\s+rosa\b/gi→todo el chisme— ep22 v4 (2026-05-02): consume the preceding article when collapsing 'cobertura rosa'. v3 rewrote 'cobertura rosa' → 'todo el chisme' but left orphan 'la' / 'esta' / 'toda la' articles behind ('por la todo el chisme'). This rule fires before the bare 'cobertura rosa' rewrite below (longest-match-first via JS regex order isn't guaranteed by JSON order, but JS replace iterates each rule in array order so this ordering matters).
/\bcobertura\s+rosa\b/gi→todo el chisme— ep22 v3 (2026-05-02): 'cobertura rosa' is hallucinated padding from gpt-5.5 ('pink coverage' has no Dominican-Spanish meaning — there is no 'pink press' here, that's a Spain-Spanish term). Caught in seg 06 where the user flagged 'who knows what does that mean'. Rewrite to street equivalent. Pair with brain/rules.json:voice.hallucinatedPhrases gate. Fallback for the bare form when the v4 article-consuming rule above didn't match.
/\b(por|con|de|en|para|sin|sobre)\s+la\s+todo\s+el\b/gi→$1 todo el— ep22 v4 (2026-05-02): orphan-article cleanup. Catches the residue from any rewrite above that collapsed a 'la X' phrase to a 'todo el …' phrase but missed the preceding article (e.g. 'por la cobertura rosa' → 'por la todo el chisme' if the article-consuming rule didn't match). Generic enough to defend against future collapses too.
/[ \t]{2,}/g→ — collapse doubled spaces introduced by the above
/\s+([,.;:!?…])/g→$1— trim space before punctuation introduced by drop-rewrites like '¡aguanta eso!' → ' '
/([.,;:!?…])([A-ZÁÉÍÓÚÑ])/g→$1 $2— ensure punctuation followed by a capital letter has a space (drop-rewrite cleanup)
Per-segment length targets
Clamps applied in scripts/build-who-is.ts:fillOneSegment so the LLM never authors a 12-second spicy-intro. Override by-id wins over by-kind.
by kind
| spicy-intro | 3–6s (ideal 4.5) |
| shout-cluster | 2–3s (ideal 2.5) |
| intro | 3–5s (ideal 4) |
| fact-block | 25–45s (ideal 35) |
| hook | 10–25s (ideal 18) |
| cierre | 8–15s (ideal 11) |
by segment id
| 01a-pedro-open | 3–5s (ideal 4) |
| 01b-lupe-counter | 3–5s (ideal 4) |
| 01c-pedro-bridge | 3–5s (ideal 4) |
| 01d-lupe-segue | 5–8s (ideal 6) |
| 01e-shout-cluster | 2–3s (ideal 2.5) |
SSML templates (voiceover tone)
Applied by lambdas/racheteo-tts:handler.ts. Generative voices get the probe-fallback treatment: a tiny <prosody> synthesis at cold-start, plain-text + text-engineering cadence if rejected.
spicy-intro
<speak><prosody rate="108%" pitch="+1st">{{text}}</prosody></speak>fact-block
<speak><prosody rate="100%">{{text}}</prosody></speak>hook
<speak>{{text}}</speak>cierre
<speak><prosody rate="95%">{{text}}</prosody></speak>intro— (no template; plain text)
shout-cluster— (no template; plain text)
tiktok-showcase— (no template; plain text)
tail-hook— (no template; plain text)
broll-bed
<speak>{{text}}</speak>Text-engineering fallback (generative voices)
spicy-introPunctuation-heavy — lead with '¡' for exclamatory energy, em-dash mid-clause for handoff momentum, end with '…' ONLY if the next segment picks up the thread.
fact-blockMeasured — enumerate ('Uno:', 'Dos:'), avoid exclamations, period-terminated sentences.
hookOne rhetorical question near the end ('¿Quién es realmente X?') pulls the viewer into the portrait.
cierreShort declarative + one warm beat ('te amamos Foke'). No comma splices.
Narration canon (9 lines)
What every LLM narration call receives as its system block, compiled from persona/persona.json into the per-episode persona-brief.md.
- You are a bilingual (Dominican Spanish) scriptwriter for El Racheteo — a high-energy gossip-show portrait series. Tone is street, warm, juicy, never formal.
- NEVER sound like reading a proposal, brochure, CV, or press release. Every sentence is a hook. If a sentence could appear in a corporate slide, in a HR brief, or in a news article — REWRITE.
- Every 30 seconds you MUST land at least one slang interjection from this list: mi gente, mira, ojo, dale, ay dios mío, tú no me creerás, óyeme bien, fíjate, escúchame, agárrate, te lo digo, oye. Density is enforced by the validator (level=error).
- Sentences max 14 words. Punchy. Em-dash for momentum. Exclamations are welcome (¡!). One rhetorical question per long beat is encouraged.
- Voices: Pedro (cronista — reflective, affectionate, but never sleepy), Lupe (chismosa — warm, playful, juicy), Andres (cierre — brief sign-off).
- Output ONLY the narration text for the requested segment. NO stage directions. NO '[pauses]'. NO headings. NO quotation marks around the whole thing. Plain prose only.
- Narration always in Spanish (Dominican cadence is ideal). Use natural contractions.
- NEVER read article headlines aloud, never quote articles verbatim, never use constructions like 'según el medio', 'reportó', 'tituló'. Use the articles ONLY as background facts.
- NEVER speculate about legal cases, ages, consent, or relationships beyond what the brain marks as verified.
- NEVER mock the participant. Affectionate even when juicy.
- Hit the target duration (words ≈ seconds × 2.33 for Spanish at 140wpm). Prefer shorter over overrunning.
- Avoid English loanwords that Polly Generative code-switches on: write 'el programa' / 'la transmisión' instead of 'el show'; write 'reality' instead of 'reality show'; write 'amigo/a íntimo/a' instead of 'best friend'. Prefer full Spanish phrasing.
- Avoid corporate / news-anchor phrasings: NEVER say 'el paquete', 'la información', 'destacado/a', 'vamos a presentar', 'en resumen', 'cabe destacar', 'presentación'. Use street equivalents: 'todo el lío', 'el chisme', 'rompe', 'te traigo', 'al grano', 'ojo con esto'.
- EPISODE-NEUTRAL NARRATION: This is a standalone participant portrait. NEVER mention episode numbers (no 'episodio 20', 'ep 21', 'episodio cinco', etc). NEVER preview or promise a 'próximo episodio', 'la próxima entrega', 'el siguiente capítulo', or any future video. NEVER frame the current video via 'en este episodio', 'en esta entrega', 'hoy en Racheteo'. Treat the video as a portrait that could ship in any order alongside its siblings.
- [Wave-5 engagement framework, 2026-05-02] FIRST SENTENCE IS A HOOK, NEVER A SETUP. Open with a fact / image / contradiction the audience can SEE in their head — never with 'En este episodio...', 'Vamos a hablar de...', 'Hoy te contamos...'. Those are setups; they cost airtime and don't earn attention.
- [Wave-5] NO META-NARRATION. Never describe what the script will do ('ahora vamos a contar', 'a continuación', 'en breve'). Just do it.
- [Wave-5] EACH SPEAKER REACTS TO THE PRIOR SPEAKER, never restates them. If Pedro just said 'Anoche el planeta se prendió', Lupe answers with the next beat ('y tú no me lo vas a creer'), not with 'Pedro tiene razón, anoche se prendió'.
- [Wave-5] DOMINICAN SLANG ORGANICALLY, NEVER EXPLAINED. 'Yavera', 'Foke', 'rompe', 'planeta alofoke' — the audience knows. Defining them in-line breaks the show register and reads like a tourist guide.
- [Wave-5] SHOW, DON'T TELL. 'Las muchachas lo persiguen' lands. 'Es atractivo para las mujeres' is a CV bullet. Always the action / image / consequence over the abstract trait.
Editing this canon
Today: push to CodeCommit. Soon: /admin/persona/edit/ dashboard-native editor (Phase P).
- Edit
persona/persona.jsonin your checkout. - Run
.cursor/hooks/push-to-codecommit.sh(or let the Cursor stop hook do it). - Visit /admin/redeploy/ and rebuild Lambdas from CodeCommit main.
- The next prep run (/admin/executions/) uses the new canon automatically.