Poshmark Growth Week

One Job Per Bot

Somu · Somasundaram Mahesh
8 October 2026 · Poshmark India Office · Chetpet
Thursday · the hands-on session
Fix what broke. Then a bot, a cut, and a model.
Bring a laptop

A job becomes a bot.

An edit becomes a transcript.

A shot becomes a model call.

Today

  • FixThe skills you set up yesterday. What broke, we fix. Bring a laptop.
  • Grok BotsHow people run one bot per job, on one shared computer.
  • EditCut a video by its transcript, and set that up.
  • GenerateWhich model for which job. The multi-model side.

Three parts. Every part ends in something you can run.

Part one

Grok Bots

Using a Grok Bot feels like texting a colleague.

They have a real desktop.

That desktop stays.

What it is

  • A BotAn AI teammate with a name, a job, and context that compounds over time.
  • A computerEvery Bot works on a persistent cloud computer. A browser, a filesystem, a terminal.
  • One computerShared by all your Bots. Cookies, sessions, and files are shared.
  • Not a draftTasks finish in your real tools, instead of as chat drafts.

docs.x.ai, Grok Bot overview and computer-and-apps.

The door

  • Two appsGrok is the chat app. Grok Bot is the agent with the computer. Different apps.
  • AccessEvery paid individual Cursor plan, Cursor Teams, or a linked SuperGrok plan.
  • WhereA desktop app on macOS, Windows, and Linux. A companion app on iOS and Android.

cursor.com/help/grok-bot and docs.x.ai/grok-bot, get-started.

Standing jobs, not questions

  • The filterA job goes to a bot when it happens on a schedule, produces something you check later, and does not need your laptop.
  • One solopreneur8 agents running 11 routines across 15+ tools.
  • One operatorAbout 30 agents. A chief-of-staff bot, a family agent, engineering bots.
  • The makersThe Grok Bot team itself runs 14 bots for design, engineering, and daily life.

christopheralarcon.com, justinbutlion, How I AI on chatprd.ai, codemypixel.

The receipts

  • Grok ShipA software factory inside Grok Bot. Hundreds of issues and pull requests a day.
  • Three bots overnightA product manager, a lead engineer, a review engineer. 52 pull requests in 12 hours.
  • Bots that run agentsThey create Cursor cloud agents, read transcripts, review proofs, and send follow-ups.
  • Her number2,000 pull requests a month. Self-reported. Their rate, not our target.

grokbot.dev, Harry Munro on LinkedIn, x.ai/bot/guides, Lauren Tan.

Part two

Edit Video

Edit the text.

The video follows.

Transcript first

  • DescriptEdit video like a doc. Change the transcript, the media updates. No timeline.
  • auto-editorA command-line first pass. It cuts the dead space by audio loudness.
  • Whisper pipelinesThey remove silences, filler words, and restarts from talking-head video.
  • Word boundariesOne skill cuts on Montreal Forced Aligner boundaries, 10 to 20 ms. It keeps the laughs.

help.descript.com, WyattBlue/auto-editor, timkulbaev/ai-video-editor, cut-video.

Say it in a sentence

  • A skillEdit video in natural language. Trim, jump cuts, captions, overlays, speed. ffmpeg and Whisper underneath.
  • A pipelineAn 8-step automated pipeline built as skills. Silence, zoom, color, audio mastering, burned-in captions.
  • Both endsDescript at the surface, or ffmpeg underneath. The transcript is the interface.

Surojit16/video-editing-skill, chrislema/videoeditor.

Set it up today

  1. ffmpeg and uv. The two prerequisites.
  2. The tool. npx skills add WyattBlue/auto-editor.
  3. The loudness pass. auto-editor path/to/your/video.mp4. Dead space gone.
  4. The transcript. auto-editor whisper video.mp4 ggml-medium.en.bin --format srt -o video.srt. The second argument is the model file.
  5. The cut. auto-editor video.mp4 --edit subtitle. It reads video.srt.

WyattBlue/auto-editor, skills/auto-editor-transcribe. The uv and ffmpeg gist.

Part three

Generate Video

No single model does the whole job.

A pipeline has a model per step.

The field, today

  • Veo 3.18-second videos, 720p to 4K, with natively generated audio. In the Gemini API.
  • Sora 2OpenAI's flagship video and audio model. Synced dialogue, sound effects. September 2025.
  • The endThe Sora app closed April 26, 2026. The API ended September 24, 2026. No replacement named.
  • The fieldWan 3.0, Seedance, MiniMax H3, Kling 3.0, FLUX 3, grok-imagine-video.

ai.google.dev, openai.com, help.openai.com, Artificial Analysis.

Quality, priced

  • The voteBlind pairwise human preference. Video with audio, from a text prompt.
  • The topWan 3.0 at 1156. Dreamina Seedance 2.5 at 1143. MiniMax H3 at 1137.
  • The priceMiniMax H3, $4.80 a minute. Veo 3.1, $24.00 a minute.
  • The catchVeo 3.1 ranks 18th on preference Elo. One axis, not the verdict.

Artificial Analysis, AA-Video-T2V v2.0, retrieved 8 October 2026.

Which model for which job

  • Cinematic realismSora 2's API ended September 24, 2026. Veo 3.1 owns the cinematic lane.
  • Image to videoKling keeps showing up in the image-to-video shootouts.
  • Editing plus generationRunway Gen-4 with timeline tooling, when editing matters as much as generation.
  • All-rounder and valueVeo 3.1, prompt adherence, native audio, 4K. Kling 3.0, value to quality, multi-shot storyboards.

motionvid.ai, toolcenter.ai, frankx.ai, 2026 comparisons.

The pipeline

  1. A model writes the script.
  2. A TTS voice speaks it.
  3. An image model draws the frames.
  4. An image-to-video model animates them.
  5. BGM and assembly land the MP4.

counter-eng/ai-video-factory. LLM, TTS, AI images, img2vid, BGM, assembly.

Who wires it

  • Eleven skillsOne job each. Research, script, storyboard, voice-over, images, timeline assembly.
  • One Google lineGemini writes. Imagen draws. Veo animates. Story to movie.
  • Multi-APIOpenAI, Gemini, and ElevenLabs in one orchestration. Sora 2's API ended September 24, 2026.
  • LocalThe whole pipeline runs on your own GPU.

chenyujinchj/video-agent-skills, SDSmirnov/AI-Story-To-Movie, wonderstone843, counter-eng.

The lesson

Do not marry the model.

  • The fallThe Sora app closed April 26, 2026. The API ended September 24, 2026.
  • The gapNo replacement was named. An endpoint migration, not a model-name swap.
  • The pipelineA model per step swaps one step. A single-model stack rewrites itself.

help.openai.com, aistackcurrent.com, September 24, 2026.

Sources · the bots

  • Grok Botdocs.x.ai/grok-bot. x.ai/bot/guides. grokbot.dev. cursor.com/help/grok-bot.
  • The fleetsjustinbutlion.substack.com. chatprd.ai/how-i-ai. christopheralarcon.com. codemypixel.com. Harry Munro on LinkedIn.

Retrieved with Exa, 8 October 2026.

Sources · the video

  • Editinghelp.descript.com. auto-editor.com. timkulbaev/ai-video-editor. cut-video. Surojit16/video-editing-skill. chrislema/videoeditor.
  • Generationai.google.dev/gemini-api/docs/veo. openai.com/index/sora-2. help.openai.com. artificialanalysis.ai. motionvid.ai. toolcenter.ai. frankx.ai.
  • Pipelinescounter-eng/ai-video-factory. chenyujinchj/video-agent-skills. SDSmirnov/AI-Story-To-Movie. wonderstone843/genai-video-pipeline.

Retrieved with Exa, 8 October 2026.

One bot, one job.

One model, one step.

The pipeline is what you keep.

Leave with this

  • FixThe skill that broke yesterday. One repo. One verify run.
  • BotOne standing job. Named, scheduled, checked later.
  • CutOne video, edited by its transcript.
  • PickBy the lane and the price, not the name.

Bring a laptop. The rest is typing.