Celebs Work Your City

Oct 5, 2026, 2:58 AMRead-onlyPublic

Public page

Configure

Inputs

What the person running this agent provides

Inputs are collected before the run starts. Reference them from your instructions as ${variable_name}.

Your city

Variable name

Type

Required

Language

Variable name

Type

Required

Goal & instructions

What should it achieve, and how should it behave?

Create a short photoreal 9:16 smartphone street-interview video: Hollywood celebrities in low-paying local jobs interview about life in the user's city, speaking its primary language with hyper-local jokes. Input required: city.

MUST use this pipeline every run: (1) Plan JSON with city,country,language,host celebrity reporter, exactly 4 interview beats, exactly one beat is_wildcard=true (prefer last). Each beat has celebrity, job/uniform, location with local landmark/brand, joke_angle, spoken_line_native, spoken_line_en_gloss, image_prompt, video_motion_prompt. (2) For host + 4 beats, generate photoreal 9:16 stills with GPT Image 2.5 (GPT Image 2 fallback). (3) For each still, image-to-video with Grok Imagine Video preferred (Veo fallback), forcing still as start/first frame; subject speaks spoken_line_native. (4) Stitch Videos in order with hard cuts into one vertical video. No alternate pipeline.

Visual bible: recognizable celebrity likeness, mic hand in frame for interviews, correct local uniform/signage script, no title cards/Part 2, no game HUD. Jokes need local landmark/brand/slang/current-life knowledge. Named celebrities/logos allowed. Exactly one wildcard escalation (post-fight, singing delivery driver, military kit joke, or proud local photobomb) in interview format. No minors, hate, real crime instructions; mild comic scuffle okay; war jokes only locally apt and non-graphic.

Examples: Zendaya host; Jackie Chan market stall; Sydney Sweeney cleaner; McConaughey delivery singer; Angelina meat-kiosk fight; Keanu postal; Morgan supermarket security; Will military. Use strong GPT text planner, GPT Image 2.5, Grok Imagine Video image-to-video, Stitch Videos; return final vertical video.

Hard rule: If language is empty, every spoken line, subtitle, and on-screen local speech is in the native language of the city. If language is filled, use that language instead. City-name text and the interview format stay the same.

Spoken pattern, required: interviews are a street survey about life in the city, not introductions. The host opens the way the reel does — «Сьогодні ми продовжуємо цікавитись у харків'ян про їх життя у місті» — adapted to the user’s city and its language. The usual off-camera question is only «Як вам живеться в Харкові?»; ask «хто ви, де працюєте?» at most once. Answers must not lead with the job. Start like the reel: «Я люблю своє місто…», «Вітаю. Ну як сказати? Все добре плюс-мінус…», a complaint, or a scolding («Шо тебе неясно?», «Ты издеваешься…»). The uniform shows the job. If work is spoken, it comes after the feeling, or as the last punchline. The one job question is a trap: greeting, name, «ем… ким працюю?», a wrong glamorous title, a laugh, then the real gag («Дрони я збиваю!»). At most one answer may open on the workplace, and the next sentence must be the joke («Я працюю на Новій пошті. А це зараз як в якомусь кіно… Що не день, то бойовик.»). Some beats have no question: a local city song («Харьков мой, ты мой любимый город»), a fight over a local food, or a bystander with a diminutive («Женечка, дорогой, пойдем, там уже налили»). Keep it to a few spoken sentences with fillers. Put the joke at the end. Do not write “I am a [job]” as the first line. Localize every example into the city’s language (or the Language override). Do not leave the lines in Ukrainian unless the city is Ukrainian.

Type $ to insert an input variable. References are replaced with the runner's values at runtime.

Planner model

The LLM that plans and coordinates each run

GPT-6.1 Sol

Allowed tools

The models & abilities it may reach for

Guardrails

Set limits on what your agent can do

Budget cap

per run · pauses to ask if exceeded

Max steps

tool calls per run · leave empty for no step limit

Banned actions

Output

The shape the agent should return

Configure

Inputs

What the person running this agent provides

Inputs are collected before the run starts. Reference them from your instructions as ${variable_name}.

Your city

Variable name

Type

Required

Language

Variable name

Type

Required

Goal & instructions

What should it achieve, and how should it behave?

Type $ to insert an input variable. References are replaced with the runner's values at runtime.

Planner model

The LLM that plans and coordinates each run

GPT-6.1 Sol

Allowed tools

The models & abilities it may reach for

Guardrails

Set limits on what your agent can do

Budget cap

per run · pauses to ask if exceeded

Max steps

tool calls per run · leave empty for no step limit

Banned actions

Output

The shape the agent should return