Most podcasts talk about AI tools. @thegodmodepod EP36 runs one live on itself At 1:08:30 Gannon Breslin points the Snowball MCP in Claude at the God Mode TikTok, pulls in 187 posts and asks what works The answer they didn't expect: TikTok files a lot of their AI videos under games. His guess is the tier list format, the same one streamers use Worth it if you post short video and keep guessing why reach stalls. TikTok weighs every video across 41 content categories, not one youtu.be/HawQxKYw9zk

An AI agent is not a model problem anymore It is an org chart problem Most setups start with one generalist Bot doing everything The Grok Bot guide lists that as mistake number one One Chief of Staff Bot routes the work and checks the result Specialists get a domain you can name in three words Anything wider is a skill, not another Bot The catch nobody reads twice Every Bot shares one cloud computer Same files, same cookies, same logins Separate Bots are not a security boundary Send, publish, pay, delete, sign, trade Those six always wait for a human Background approvals expire in about 10 minutes and the action never runs Spinning up Bots is the easy part The org chart and the gate list are the real work Full breakdown below

Frame by frame on my @Kling_ai 4.0 clip The car blinked when it woke up, and the tiny driver inside blinked with it. I never mentioned a driver Off the table, the camera dropped with the car in one 5 second shot, no cut On landing it squinted like it felt the hit, then opened its eyes Adaptive ratio kept my square source, so the clip came out 940 by 940 One fix: 66 free credits and a 50 credit clip leave no second try. A cheap low res preview pass before the full render would help

One image, one prompt, @Kling_ai 4.0 I gave it the TestDrive car and asked: wake up, blink, drive off the table toward the camera What it got right It drove off the edge and the camera followed it down to the floor The face, orange bumper and little driver survived the fall unchanged It added sound on its own What it cost One 5 second 720p clip took 50 of my 66 free credits Ready in under 6 minutes in the free queue Free account math: you get one Kling 4.0 shot, so write the whole scene in that prompt

The edit already takes orders in plain English Now the footage does too A plugin in Claude Code cuts this talking head from words Subtitles, music, motion graphics, B-roll, auto zooms Then a timeline and an XML export Since October 15 Kling 4.0 answers the same terminal Through MCP and the CLI, before its own API Clips of 3 to 30 seconds, prompts up to 8,000 tokens Not every web option made it into MCP The old way in was a listed $700 API pack, models up to 3.0 The new way in is a plan listed at $32.56 and a browser login Footage and cut can now share one session Whoever writes the instructions runs the whole studio Full breakdown below

I tested @prelint_ai on a PR built to break the rules Private test repo, one SPEC.md with 4 product rules: 740 char posts, one post per campaign, never store emails, rewards in USD with 2 decimals Then a PR that quietly broke all 4 What happened Review landed 75 seconds after I opened the PR All 4 caught, inline on the exact lines, verdict Reconsider It even caught the subtle one: Math.round turns 12.50 into "13 USD" The part I didn't expect A 5th finding I never planted: anyone can type an X handle without proving it's theirs One fix: the landing says a 20 second review, my first one took 75

18 reviews on @testdriveso since Sep 29, here is my honest take What works One product, one task, one tag per brief. I install and run the thing before I write The draft check flags a missing tag before I post Rewards shows views, likes, replies and bookmarks for every post I recorded What I noticed Every pool is fixed and split across everyone who joins $400 across 21 reviews and $500 across 92 are very different drives, and my views barely change that One fix first: show an estimated payout per review next to the review count before I reserve

I gave Claude Code its own computer with @trycua Cua Spaces on a MacBook Air: connect Claude Code, spin up an Ubuntu Space, a separate desktop inside the Mac, so the agent never touches my files or mouse One task: open a browser, go to testdrive.so/campaigns, save the live campaigns with reward pools to campaigns.txt What happened 40 seconds, 6 Cua calls Chromium opened in the Space and loaded the page on screen 5 live campaigns saved to campaigns.txt The part I liked It read the page text instead of guessing from pixels, and ignored the "sign up and post" instructions it found there One fix: open the saved file on the Space screen so I can check it without asking

I sketched my Excalidraw review as a diagram, without drawing a single box Opened excalidraw.com, no account More tools, Mermaid to Excalidraw, typed 7 lines of flowchart text The preview redraws as you type, Insert puts it on the canvas in the hand drawn style What worked Every box and arrow stays editable, I colored the good and fix branches after It kept my last code when I reopened the dialog What I'd fix Only flowchart, sequence, class and ER convert. My pie chart came in as a flat image: no sketch look, no color controls, and it landed on top of my diagram @excalidraw

Video from Claude Code is not a model question It is an access question This clip runs the whole loop from a terminal One photo in a folder, one message The agent picks a model, checks the price, saves the file Kling 4.0 reached that loop before its own API did MCP and CLI since October 15 The developer API is still coming soon The listed API route opens with a $700 pack and stops at 3.0 The agent route opens with a plan listed at $32.56 One command adds the server to Claude Code A browser login hands over three scopes Create generations, read them, check credits Access is a login now The rules handed to the agent are what ships Full breakdown below

Free model, no API key, one prompt @OpenBot_ still finished a real task Windows app, OpenCode free model: DeepSeek V4.1 Flash One prompt made Campaign Scout: open testdrive.so, list live campaigns, save a file It started on its own, about 2 minutes later campaigns.md was there 7 of 7 campaigns, every review count matched the site A second run did it in about a minute, adding tags and post limits The rough part: while it worked, the chat showed raw half-sentence reasoning What I would add: a clean status line instead Rewarded TestDrive review

I ran the TestDrive Test Track, the first public run Join a Test Drive, write the review, publish it from the page What was clear Joined with my verified X account and read the brief first The draft check marks each requirement Met or Not yet as you type Reserve holds the exact text, then Copy the text and post on X Where I had to stop and think Pasting into the draft once dropped my blank lines The post link still goes back to the Test Drive page by hand One fix: keep blank lines when pasting into the draft Joined a rewarded Test Drive from @testdriveso

The best Grok Bot feature is hidden on purpose Skills got buried so you would just ask in plain words Per the team, 99% of automations start as a plain request Type / and the reusable skills show up anyway The recorder turns one run into a skill Do the task once in the bot browser, up to 10 minutes One operator recorded a weekly report that way It has run unattended ever since The marketplace already holds 54 free bots Haggle Bot and Projects Manager among them Teslaconomics made $723 in two weeks from shared templates Paid through X Money, invite only The prompt is not the asset anymore The recorded skill is Full breakdown below

A free AI visibility scan is a great idea Until your whole brand lives on someone else's domain I ran Agentsy Report on my X profile, no signup, about 2 minutes It dropped /Skaly__Bull and scanned X itself Result: X (formerly Twitter), named in 3 of 13 AI answers, cited 0% The checks look real: AI crawler access, 11 of 14 basics flagged They just describe X, not me What I would add: catch profile URLs and ask for the brand name Plenty of creators only have a profile Rewarded TestDrive review

One prompt and zero assets built this race graphic The agent wrote code, not a single frame by hand Claude Code went into plan mode Wrote the components, rendered Albert Park Motion graphics were already a terminal job Generated footage is the part that just moved Kling 4.0 takes orders from Claude Code through MCP On the agent route, 3 to 30 seconds at 720p or 1080p On the listings the old API route opens at $700 And it still has no 4.0 A 30-second 1080p clip on Pro works out to about $4.88 on paper Code draws the graphics, credits pay for the footage The storyboard is where the work lives now Full breakdown below

Most agents start every chat from scratch Hermes Agent writes itself a skill instead From @NousResearch, free and MIT licensed After a hard task it saves the steps as a skill, then refines it One gateway for Telegram, Discord, Slack and CLI, plus cron Why @testdriveso should test it 250K+ GitHub stars say people install it The real question: is run 10 actually better than run 1? The test I would run Same weekly report task, 10 runs, time and tokens per run Rewarded TestDrive scouting

A cat and an otter fight with swords in this test The interesting part is who placed the order Since October 15 Kling 4.0 takes orders from Claude Code Through MCP, on a regular Kling plan The 4.0 API is still marked coming soon The character reference plate is the real feature Up to 15 references per task, one shared ceiling Up to 10 images and up to 5 videos inside it Up to 7 subjects, plus voice references Programmatic Kling used to mean a listed $700 API pack That pack still stops at 3.0 Now the entry is a plan listed at $32.56 a month The reel is the easy part now Choosing what fills those 15 slots is the work Full breakdown below

The film crew in this clip does not exist Neither does an API for the model that shot it Kling 4.0 went live on October 15 On the web, through MCP and through the CLI The developer API is still marked coming soon So is the 10-bit HDR flashing on screen here Scripting Kling used to start with a listed $700 API pack Now it starts with a plan listed at $32.56 a month One command adds it to Claude Code A browser login replaces the key Then clips of 3 to 30 seconds in a single pass Up to 10 keyframes or up to 15 references The crew behind the camera is a prompt now What still sells is the brief behind it Full breakdown below

Most coding agents ship every feature on day one Pi ships almost none, on purpose Pi from @pidotdev: a free, MIT licensed terminal coding agent No sub-agents, no plan mode, no permission popups by default You add what you need as TypeScript extensions Why @testdriveso should test it Does a minimal harness beat all-in-one agents once it is set up? The test I would run One bug, Pi with two extensions vs a full featured agent Compare tokens, tries and setup time Rewarded TestDrive scouting