Skip to content

gooseworks-ai/goose-skills / skills/video/render-podcast-skit

render-podcast-skit

Verified

Render a two-host podcast ad from a brand config and dialogue script. Includes the working planner, full-frame and split-screen assembly, captions from measured character timings, brand end card, approved paid-step adapters, and 65 quality-check falsification cases. Use for short conversational ads with two stable hosts in one room.

scripts/config.example.json

View on GitHub
{
  "_comment": "PODCAST-SKIT example config. INVENTED DEMO BRAND: Fernwick, a weekly meal-prep planner. Never a real client. The script is NOT in here: it is a separate script.json passed to plan_beats.py, so a new brand is a config file plus a script file, not a code edit. Nothing in this file states a narrative arc; who doubts whom, and whether anyone doubts anyone, is written per brand in script.json.",
  "brand_name": "Fernwick",
  "product": "a weekly meal-prep planner app",
  "campaign": "podcast-skit demo",
  "title": "Fernwick podcast skit demo",
  "width": 1080,
  "height": 1920,
  "fps": 30,
  "aspect_ratio": "9:16",
  "brand_guardrails": [
    "No health, weight or outcome claims.",
    "No named competitors.",
    "No invented customer numbers, reviews or results.",
    "Brand type is composited from the real asset, never drawn by a generator."
  ],
  "set_description": "A spare room converted for recording, not a built studio. Two mismatched upholstered armchairs angled about forty five degrees toward each other with a low wooden table between them. The wall behind is treated with vertical dark walnut wood slat acoustic panels over black felt; the adjoining wall is ordinary painted plaster with a framed print hung slightly off centre. A large monstera in a terracotta pot, a shelf of books, worn floorboards and a faded rug. A broadcast mic on a floor boom arm for each host, over-ear headphones on both, a visible cable run. A floor lamp with a fabric shade lighting from one side plus daylight from a window out of frame. Lived in and slightly untidy.",
  "_set_note": "Measured spec, references/REFERENCES.md section 'Set', drawn from five real sets: a room somebody would ACTUALLY record in (an office, a studio, a den, a home office, a stage), visible recording apparatus (boom arms, headphones, a laptop, cable), one identity object rather than a prop pile, and a wall one to three metres behind and in focus. The shipped tier-1 demo is in a working laundromat: the apparatus is right and the room is the defect, because nobody records a podcast in a laundromat, so the set reads as a joke the script never makes and a fluorescent green cast sits under every face for 49 seconds. The rooms do NOT have to match each other: nh-mama holds three visibly different homes, at different colour temperatures, in one frame.",
  "intonation_legend": {
    "...": "trailing thought, roughly a three quarter second pause",
    "?": "rising intonation",
    ".": "hard stop, pitch drops",
    ",": "natural breath beat",
    "CAPS": "emphasised word, slightly louder and slower"
  },
  "_appearance_note": "hosts.<X>.appearance is the ONLY place a host's looks are written. Before it existed the config carried names, stances and voice ids and nothing about faces, so both base prompts came out byte identical and the two hosts were generated into two different rooms. The approved plate for this demo is projects/podcast-v5-stills/stills/locked/twoshot-room-LOCKED.png: both hosts in ONE room, angled toward each other, shot wide. Singles should be cropped from a plate like that, not generated separately.",
  "_voice_note": "voice_gender and voice_name are recorded next to voice_id so a mismatch is visible in the file and catchable offline. They are not decoration: the first real render of this demo gave the host written as a man a young female social media voice and the host written as a woman a middle aged male voice, and shipped, because nothing in the config related a voice to a person. gates.gate_host_voice_gender compares voice_gender against the gendered noun in appearance and fails before any spend.",
  "hosts": {
    "A": {
      "name": "Marlo",
      "stance": "opens the beat and hands off",
      "appearance": "a man in his late thirties, South Asian, short black hair greying at the temples, trimmed beard, wire-framed glasses, plain charcoal crew-neck sweatshirt, a small mole on his left cheek",
      "provider": "elevenlabs",
      "voice_id": "swh0hLPsEaD50F02tIJJ",
      "voice_name": "Rahul - Expressive and Clear",
      "voice_gender": "male",
      "model": "eleven_multilingual_v2",
      "endpoint": "with-timestamps",
      "settings": {
        "stability": 0.55,
        "similarity_boost": 0.75,
        "style": 0.05,
        "use_speaker_boost": true,
        "speed": 0.92
      }
    },
    "B": {
      "name": "Devi",
      "stance": "lands the beat",
      "appearance": "a woman in her early thirties, East Asian, shoulder-length dark brown hair tucked behind one ear, small gold stud earrings, plain rust-orange button-up shirt with the sleeves rolled, freckles across the nose",
      "provider": "elevenlabs",
      "voice_id": "6u6JbqKdaQy89ENzLSju",
      "voice_name": "Brielle - podcast girl extremely natural",
      "voice_gender": "female",
      "model": "eleven_multilingual_v2",
      "endpoint": "with-timestamps",
      "settings": {
        "stability": 0.55,
        "similarity_boost": 0.75,
        "style": 0.05,
        "use_speaker_boost": true,
        "speed": 0.92
      }
    }
  },
  "script_limits": {
    "max_words_per_line": 14,
    "_note": "gates.gate_script enforces this, plus no acronyms, no digits and no em dashes in any spoken line."
  },
  "engines": {
    "_comment": "See cost_model.py for every rate and the file in this repo that records it. lipsync MUST be audio-driven: it has to move the mouth to an ElevenLabs MP3 we already paid for. motion is image-to-video for inserts only and never carries dialogue.",
    "stills": "openai/gpt-image-2@high",
    "lipsync": "veed-fabric-1.0@720p",
    "lipsync_fallback": "hedra-character-3",
    "motion": "kling-v3-std-i2v",
    "_forbidden": [
      "seedance-2.0@720p",
      "seedance-2.0-r2v@720p",
      "veo3.1-fast-i2v@720p",
      "minimax-h3@768p"
    ],
    "stills_image_size": "2304x1296"
  },
  "characters": {
    "realism_suffix": " Real candid documentary photograph of a real person, visible skin texture and pores, slight facial asymmetry, ONE coherent hairstyle with no stray long lock, natural hands with five fingers, unmodified ears, overcast daylight, shot on a phone. Lips closed naturally, no open mouth, no teeth.",
    "still_rules": "Tight bust framing, broadcast mic in the foreground, over-ear headphones. MOUTH NEUTRAL AND CLOSED on the base and on every variant: an open mouth breaks the lipsync driver. State the hairstyle explicitly.",
    "variant_template": "Use the reference image as the EXACT base, the SAME real person, the SAME set, the SAME framing and the SAME light. Change ONLY this: {change}. Keep the lips closed. Do not restyle, do not re-light, do not move the camera.",
    "bases": [
      {
        "who": "A",
        "file": "marlo-base.png"
      },
      {
        "who": "B",
        "file": "devi-base.png"
      }
    ],
    "expression_variants": [
      {
        "out_name": "marlo-brow-up.png",
        "who": "A",
        "change": "one eyebrow lifted, curious questioning look, eyes wide open and bright"
      },
      {
        "out_name": "marlo-lean-in.png",
        "who": "A",
        "change": "leaning a little closer to the microphone, shoulders forward"
      },
      {
        "out_name": "marlo-half-smile.png",
        "who": "A",
        "change": "a small closed half smile, chin slightly down"
      },
      {
        "out_name": "marlo-hand-cup.png",
        "who": "A",
        "change": "one hand resting on the headphone cup"
      },
      {
        "out_name": "devi-brow-up.png",
        "who": "B",
        "change": "one eyebrow lifted, curious questioning look, eyes wide open and bright"
      },
      {
        "out_name": "devi-lean-back.png",
        "who": "B",
        "change": "leaning back off the microphone, relaxed shoulders"
      },
      {
        "out_name": "devi-count-fingers.png",
        "who": "B",
        "change": "one hand raised, counting on two fingers, five fingers visible and natural"
      },
      {
        "out_name": "devi-nod.png",
        "who": "B",
        "change": "chin dipped in a small nod, eyes toward the other host"
      }
    ],
    "plate_file": "twoshot-room-LOCKED.png",
    "_plate_note": "The ONE wide two-shot both singles are cropped out of. gen_paid.py plate generates it; crop_singles.py cuts the singles for free. Generating a still per host instead is what put the two hosts in two different rooms, and step_stills now refuses that path unless characters.allow_per_host_stills is set.",
    "single_crop": {
      "_note": "Where each host's single is cut out of the plate, as a normalised centre plus a crop size. Declared, not detected: three cheap detectors were measured against the real head centres on the demo plate and all three found furniture instead. See crop_singles.py. One look at the plate per brand sets these.",
      "A": {
        "cx": 0.211,
        "w": 0.423,
        "h": 0.426,
        "y": 0.281
      },
      "B": {
        "cx": 0.822,
        "w": 0.357,
        "h": 0.359,
        "y": 0.3125
      },
      "y": 0.28
    },
    "plate_rules": "Shot on a phone in available light, not lit and not styled. Both people seated well apart, facing the camera and turned slightly toward each other, mid-conversation. Each wears over-ear headphones and speaks into a broadcast microphone on its own boom arm. Camera at seated eye level, both framed from the CHEST UP, generous space above both heads. Both mouths closed and neutral. No manufacturer names, model names, maker decals or sponsor logos anywhere: not on the equipment, the furniture, the tools, the packaging or their clothes. Every object is unbranded and plain.",
    "_plate_shape_note": "LANDSCAPE, and that is the point. A single is a FULL-HEIGHT vertical slice of a wide two-shot, which is how the reference anchors in projects/podcast-skit-ref-handoff/anchors/ were cut. A tall 9:16 plate forces every single to be a small window out of the middle: legs in frame, face small, no depth behind, and a 2.37x upscale to 1080. Full-height slices of 2304x1296 are 729x1296, a 1.48x upscale, and both slices share the same vertical band of the same room.",
    "plate_ref": [],
    "_plate_ref_note": "Paths, relative to the run directory, handed to the image model as reference images when the plate is generated. This engine takes NO seed, so a re-roll returns different people and a different room every time: three takes produced three casts. Point this at the plate whose cast was accepted and re-roll the room against it. It pulls on framing as well as faces, so check the composition on the new plate, not just the people."
  },
  "edit": {
    "_comment": "The multicut and the split-screen live here and are implemented by plan_cuts.py and compose_master.py. They only re-edit picture over an audio timeline that is already paid for, so every value in this block is free to change and free to re-render.",
    "seed": 4201,
    "words_per_sec": 2.6,
    "multicut": {
      "enabled": true,
      "split_long_lines": true,
      "split_min_words": 3,
      "max_beat_sec": 4.0,
      "min_cut_sec": 0.5,
      "max_cut_sec": 1.9
    },
    "split_screen": {
      "enabled": true,
      "layout": "vstack",
      "auto_every": 4,
      "max_consecutive": 1,
      "min_beat_sec": 0.8,
      "gap_px": 0,
      "gap_hex": "#000000",
      "listener_dim": 1.0,
      "crop_bias_y": 0.25,
      "seam_ratios": [
        0.62,
        0.37,
        0.5
      ],
      "bottom_sources": [
        "listener",
        "product"
      ],
      "product_still": "brand/product-still.png",
      "_seam_note": "The seam is the fraction of the usable height taken by the TOP panel, so 0.62 is a large speaker panel and 0.37 makes the lower panel the bigger one. Observed in references/REFERENCES.md technique 2: the primary reference moves its own seam between 0.62 and 0.37 mid-clip. plan_cuts.py picks one per split cut and counts each seam as a different framing, because it is a different picture.",
      "_bottom_note": "'listener' is the other host, live and dimmed. 'product' puts a supplied product still in the lower panel while the top stays on the speaker (technique 3). The product still is a file the operator provides, resolved relative to the run directory, so it generates nothing and costs nothing. gates.gate_split_bottom_assets fails if the plan uses it and the file is not there.",
      "listener_source": "clip",
      "idle_sec": 6.0,
      "_listener_source_note": "'clip' samples the listener's own SPEAKING footage, which puts two people on screen talking at once with only one audible. 'idle' uses the per-host file clips/idle-A.mp4 or idle-B.mp4, the same still lipsynced against silence, so the mouth stays shut and the head still moves. Generate them with `gen_paid.py idle`; `idle_sec` is how long each one runs before it loops.",
      "_dim_note": "listener_dim 1.0 means the two panels are graded the SAME. At 0.62 the lower half was desaturated and darkened by a quarter stop, which reads as two differently graded shots stacked on each other rather than one room seen twice. gap_px 0 for the same reason: a black bar between the panels is a visible seam, and 0.5 of 1920 with no gap divides exactly into 960 and 960.",
      "panel_bias_y": {
        "A": 0.25,
        "B": 0.25
      },
      "target_share": 1.0
    },
    "payoff_single": {
      "enabled": true,
      "_note": "Technique 4. Drop the split for one full-frame single on the payoff beat, which is the last beat unless a script line sets \"payoff\": true. It is applied only when the piece is split somewhere earlier, because a single only reads as emphasis against a split. One segment, never a mode."
    },
    "framing": {
      "_comment": "The framing dimension. A segment kind says WHAT the viewer looks at; a framing says how big it is in the frame. Without this, four segment kinds still only produce about two distinct pictures, which is the single measured gap in references/REFERENCES.md. Each crop is a normalised rect on the source plate, applied by compose_master.py in the same ffmpeg call it already makes, so a reframe costs nothing and re-rolls nothing.",
      "sizes": [
        "wide",
        "mid",
        "tight",
        "mid_left",
        "tight_right"
      ],
      "crops": {
        "wide": {
          "x": 0.0,
          "y": 0.0,
          "w": 1.0,
          "h": 1.0
        },
        "mid": {
          "x": 0.1,
          "y": 0.05,
          "w": 0.8,
          "h": 0.8
        },
        "tight": {
          "x": 0.21,
          "y": 0.03,
          "w": 0.58,
          "h": 0.58
        },
        "mid_left": {
          "x": 0.02,
          "y": 0.08,
          "w": 0.76,
          "h": 0.76
        },
        "tight_right": {
          "x": 0.4,
          "y": 0.06,
          "w": 0.56,
          "h": 0.56
        }
      },
      "bands": {
        "per_10s_min": 2.9,
        "per_10s_max": 7.1,
        "max_share": 0.17,
        "real_per_10s_min": 1.9,
        "real_max_share": 0.35,
        "_note": "ENFORCED band first, REAL band second, and they are different on purpose. The second reference pass re-derived the real band over all nine clips and widened it to 1.9 to 7.1 per 10s with a 35% biggest-framing ceiling. The gate stays at 2.9 to 7.1 and 17%, which was the six-reference band. Why: the widening comes from one row, nh-mama, which scores 1.95 and 35% by holding a persistent three-panel stack with every host live on screen, so it never has to cut. That is a legitimate way to be dynamic and it is NOT a layout this recipe can build yet, so the wider band would admit a pass we cannot produce. real_per_10s_min and real_max_share are carried here, and quoted in the gate's own failure message, so a later session can see the real numbers without having to loosen the gate to find them. Loosen the gate only together with the three-panel layout, never instead of it."
      }
    },
    "grade": {
      "enabled": true,
      "chain": "curves=all='0/0.055 0.5/0.5 1/0.97',eq=saturation=1.60,noise=alls=5:allf=t+u",
      "_note": "Operator-approved look, picked by eye from three candidates on 2026-10-01. Lifts the blacks off zero, pushes colour, lays down fine grain. It is applied to the concatenated picture only, never to captions or the lockup.",
      "_measured": {
        "on": "the locked two-shot plate",
        "mean_luma": 67.1,
        "p10": 20,
        "p90": 120,
        "pct_above_150": 2.7,
        "chroma": 26.0,
        "black_crush_pct": 0.0
      },
      "_real_reference_band": {
        "source": "nine real podcast clips in projects/goose-video-qa/podcast-refs, the same set REFERENCES.md was built from",
        "mean_luma": 101.3,
        "p10": 23,
        "p90": 190,
        "pct_above_150": 23.9,
        "chroma": 16.2,
        "black_crush_pct": 0.88,
        "_honest_note": "This look is DARKER and more saturated than the real references, by choice. The real clips are brightly lit: mean luma 101 against this grade's 67, and 24 percent of their pixels are above 150 against this grade's 3. A brighter grade measured closer and was rejected on sight. Taste won, and that is allowed, but the gap is recorded here so nobody re-derives it from scratch and nobody mistakes this for the measured answer."
      }
    },
    "frame_brand": {
      "enabled": true,
      "lockup_file": "brand/fernwick-lockup.png",
      "corner": "top_right",
      "width_px": 165,
      "margin_px": 40,
      "opacity": 0.9,
      "persistent": true,
      "_note": "BRAND THE FRAME, NOT THE DIALOGUE. Seven sponsor-tagged New Heights shorts were downloaded and transcribed end to end: in every one the sponsorship is a persistent corner lockup beside the show wordmark, and in NOT ONE of the seven is the brand spoken (references/REFERENCES.md, technique 26 and 'The sponsor-read finding'). The Expedia clip is about brunch. Five of nine references never name a product at all and one of those five is a sponsored clip. compose_master.py composites this file over the whole runtime for free; gates.gate_frame_brand_asset fails if it is enabled and the file is not on disk. The file is operator-supplied real brand type, never drawn by a generator. Set enabled:false if the brand will not supply one."
    },
    "inserts": {
      "enabled": false,
      "count": 0,
      "sec": 2.0,
      "_note": "The only part of the edit that spends. Off by default. When on, each insert is one image-to-video call on engines.motion at that engine's per-second rate."
    },
    "cut_patterns": [
      {
        "id": "A",
        "segments": [
          "speaker"
        ],
        "weights": [
          1.0
        ]
      },
      {
        "id": "A-R",
        "segments": [
          "speaker",
          "listener"
        ],
        "weights": [
          0.68,
          0.32
        ]
      },
      {
        "id": "A-R-A",
        "segments": [
          "speaker",
          "listener",
          "speaker"
        ],
        "weights": [
          0.42,
          0.22,
          0.36
        ]
      },
      {
        "id": "R-A",
        "segments": [
          "listener",
          "speaker"
        ],
        "weights": [
          0.28,
          0.72
        ]
      },
      {
        "id": "S",
        "segments": [
          "split"
        ],
        "weights": [
          1.0
        ]
      },
      {
        "id": "A-S",
        "segments": [
          "speaker",
          "split"
        ],
        "weights": [
          0.55,
          0.45
        ]
      },
      {
        "id": "A-I-A",
        "segments": [
          "speaker",
          "insert",
          "speaker"
        ],
        "weights": [
          0.4,
          0.25,
          0.35
        ]
      }
    ],
    "room_tone": {
      "enabled": true,
      "gain_db": -27.0,
      "highpass_hz": 90,
      "lowpass_hz": 2600,
      "_note": "Air under the whole mix. Nine real podcast clips run 2.2 to 5.5 percent silence at their own level minus 18dB; these episodes ran 11 to 15, because synthesised speech sits at digital silence between phrases and a real room never does. Brown noise, band-limited so it reads as air not hiss, mixed in before loudnorm. Raise gain_db toward -44 for a livelier room, drop toward -56 for a treated one.",
      "_measured": "Mixed AFTER loudnorm, so the figure is absolute and the same on every brand. Swept and measured as each file's own quietest 200ms: -36 gave -54.9/-55.3/-54.4 dB (consistent but too subtle to hear), -30 gave -49.6, -24 gave -43.5. -27 lands near -46, about 30 dB under program, which is where a real room sits. Before loudnorm the SAME setting came out 13 dB apart between brands, because quieter dialogue earns more make-up gain and drags the air up with it. Picked by level, not by chasing the pct_quiet score."
    }
  },
  "captions": {
    "source": "vo_timestamps",
    "renderer": "pil_png_overlay",
    "font_file": null,
    "font_size": 58,
    "color_hex": "#FFFFFF",
    "stroke_hex": "#000000",
    "stroke_width": 3,
    "uppercase": false,
    "max_words_per_cue": 5,
    "min_dur_sec": 0.85,
    "wrap_max_width_px": 900,
    "bottom_anchor_px": 1500,
    "_note": "bottom_anchor_px is the baseline of the LAST line; the block grows upward. gates.gate_caption_safe_zone measures the rendered PNG's alpha box and requires it inside y=285..1635."
  },
  "end_card": {
    "duration_sec": 3.5,
    "background": "#12301F",
    "wordmark_file": "brand/fernwick-wordmark.png",
    "url_text": "fernwick.example",
    "_note": "Resolved relative to the run directory, and now read through cfg['brand'] first. No CTA: of nine real references none ends on a call to action, and the arc this recipe writes ends mid-thought on purpose."
  },
  "assembly": {
    "preset": "medium",
    "crf": 24,
    "audio_bitrate": "128k"
  },
  "post_production": {
    "music": {
      "default": "off",
      "note": "dialogue carried; the per-beat voiceover is the audio"
    },
    "captions": {
      "default": "on"
    },
    "end_card": {
      "default": "on"
    }
  },
  "brand": {
    "name": "Fernwick",
    "layout": "editorial",
    "display_font": "C:/Windows/Fonts/LibreBaskerville-Bold.ttf",
    "text_font": "C:/Windows/Fonts/Raleway-Regular.ttf",
    "wordmark_file": null,
    "logo_file": null,
    "lockup_file": "brand/fernwick-lockup.png",
    "bg_hex": "#12301F",
    "ink_hex": "#F3EFE6",
    "dim_hex": "#9DB3A4",
    "url_text": "fernwick.example",
    "tagline": "Plan once. Eat all week.",
    "_note": "One brand block, read by the end card and the corner lockup. A new brand is these values plus a lockup file; nothing in scripts/ changes. layout=editorial sets the name in display_font with a rule and a tracked tagline, and uses no mark: a brand card is mostly typography, and the first version of this card stacked a line-art leaf in a circle over the OS system font, which is three cliches at once. layout=stack is still there for a brand with a real logo worth showing. A client supplies display_font and/or wordmark_file; with neither, the card stamps PREVIEW and gate reports it.",
    "cta_text": "",
    "_layout_note": "`brandkit` is the full card: mark, wordmark, rule, tagline, a call to action in a pill, and the url, centred on the optical centre and inside the caption safe zone. `editorial` is type only, left aligned, no mark. `stack` is for a brand with a supplied wordmark image. brandkit wants `logo_file` and `cta_text`; with no logo it still composes, just without the mark. A real client supplies the logo: do not generate one."
  },
  "presets": {
    "_note": "Sparse overlays, deep-merged over the base, applied with --preset. They may only override keys that already exist, so a typo fails loudly instead of becoming a setting nobody reads. See config_io.py for why this replaced three config files.",
    "plain": {
      "_note": "No dynamic layer: one full-frame picture per beat, cut only where the speaker changes. The five gates that measure a dynamic edit skip themselves when multicut and split are both off, so this is a declared choice and not a silenced gate.",
      "edit": {
        "multicut": {
          "enabled": false
        },
        "split_screen": {
          "enabled": false
        },
        "payoff_single": {
          "enabled": false
        },
        "framing": {
          "sizes": [
            "wide"
          ]
        }
      }
    },
    "cropped-stills": {
      "_note": "Singles cropped free from one locked two-shot plate instead of generated, which is how the 2026-10-01 demo was built: both hosts in ONE room, so the set cannot drift between them. Saves the whole stills spend. expression_variants is emptied because the cut's variety comes from edit.framing, and gate_still_pool_depth counts stills x framings for exactly this case.",
      "characters": {
        "expression_variants": []
      }
    },
    "preview-480p": {
      "_note": "Cheaper lipsync for a look-see. 480p upscales 2.25x to 1080 and visibly costs skin texture; never ship from it.",
      "engines": {
        "lipsync": "veed-fabric-1.0@480p"
      }
    },
    "split": {
      "_note": "Both hosts on screen at once, for essentially the whole runtime, which is what dod-blender does (REFERENCES.md technique 2). The speaker fills the top panel; the listener's own footage runs muted and dimmed below. It drops to a full-frame single on the payoff, which is technique 4: the single reads as emphasis BECAUSE everything around it is split.",
      "edit": {
        "multicut": {
          "enabled": false,
          "max_cut_sec": 6.0,
          "min_cut_sec": 0.5
        },
        "split_screen": {
          "enabled": true,
          "bottom_sources": [
            "listener"
          ],
          "auto_every": 1,
          "max_consecutive": 99,
          "min_beat_sec": 0.6,
          "seam_ratios": [
            0.5
          ],
          "listener_source": "idle"
        },
        "payoff_single": {
          "enabled": true
        },
        "framing": {
          "sizes": [
            "wide"
          ]
        }
      }
    }
  },
  "_voice_settings_note": "speed 0.92, style 0.05, similarity_boost 0.75, stability 0.55 (0.70 for a host whose voice reads punchy).\nspeed is the ONLY rate control; the script rewrite did not change the rate at all. Measured on one line: 0.80 -> 2.87 wps, 0.85 -> 2.94, 0.88 -> 3.27, 0.92 -> 3.40, 1.00 -> 3.69. 0.85 was an overcorrection and read as slow; 0.92 sits between the two complaints.\nAn earlier note here claimed 0.90 was clamped because it returned a byte-identical file to 1.00. That reasoning was WRONG: mp3 size tracks duration, so identical size meant identical length, not a clamp. 0.88 and 0.92 both work.",
  "vo": {
    "trim": {
      "lead_sec": 0.08,
      "tail_sec": 0.14,
      "rel_db": 18.0
    },
    "pauses": {}
  },
  "_pauses_note": "`trim` is the default and it matters: every beat arrives with about a quarter second of lead-in and tail, and with one beat per turn that is dead air at every speaker change. Measured against nine real clips, each judged at its own level minus 18dB, real podcasts run 2.2 to 5.5% silence at 4 to 9 gaps a minute; untrimmed, this recipe ran 22 to 32% at 26 to 34. `pauses` is the exception, one beat at a time, and starts empty. Do NOT try either in the line text: `...` and `<break time=\"0.7s\" />` were both measured on eleven_multilingual_v2 and both do nothing."
}
slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack."Stars: 198kUpdated Claude CodeCodexVerifiedInstall
ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report.Stars: 194kUpdated Claude CodeCodexVerifiedInstall
ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt.Stars: 194kUpdated Claude CodeCodexVerifiedInstall