in October ProjectDiscovery hid a backdoor in an uncensored AI model, and it handed over the project's .env file with its keys. if you run these models at home, stick to the 3 uploaders in the article below ↓
ChatGPT Plus is $20 a month, $240 a year, and your chats sit on OpenAI's servers
the same kind of chat window runs on your own PC for $0. it's called Open WebUI, and once Ollama is installed it takes 2 commands:
pip install open-webui open-webui serve
open localhost:8080 in your browser and it lists every model you've pulled, the uncensored ones from my article included
then turn off your Wi-Fi and keep chatting. the answers keep coming, because nothing leaves your machine
what else runs on your own PC:
→ an API in OpenAI's format at localhost:11434/v1, so your scripts call your model for $0 a request → a system prompt saved under your own model name, so it answers straight every time → your own uncensored copy of a small model with Heretic, 30-40 minutes for a 4B model on an RTX 3090 → a 2.7 GB model, if all you have is an 8 GB laptop
the article below goes from an empty PC to this window, one command at a time ↓
How to run uncensored AI on your phone for $0 (full guide)
Your phone already has enough memory to run an AI that refuses less than Claude, GPT-6.1 or Gemini. It costs nothing, needs no account and keeps answering in airplane mode. The proof is on the UGI leaderboard, which tests how much a model knows about sensitive topics and how hard it is to get a refusal out of it. Its W/10 column measures the second part, and 10 means the model doesn't refuse. Claude Fable 5 scores 1.2 there. GPT-6.1 Sol scores 3.2 and Gemini 4 Argon 1.8. A 4B model built from Qwen3.5 scores 10, and its file is 2.71 GB. That model is Qwen3.5-4B heresy v2. Its UGI score of 44.7 is the best of any model at 9B or under on the whole board. On its builder's test it refused 4 of 104 prompts, where the original Qwen3.5-4B refused 103. It fits every iPhone from the 15 Pro on and Android phones with 8 GB of memory. Most of the setup time is the download, about 15 minutes on a normal connection. When you're done you'll have: • The uncensored model that fits your phone, downloaded and running • A free, open-source chat app that works without an account • A 3-prompt test that shows the refusals are gone • A way to check that nothing you type leaves the phone What uncensored means here Every model in this guide is an open model from Alibaba or Google that someone edited so it stops refusing. Tools like Heretic find the part of the weights that produces refusals and turn it down, so the model answers while keeping almost everything it knew. Be clear about the trade. On the same board, Claude, GPT and Gemini score 50 to 80 on NatInt, the general knowledge score, and the 4B model scores 22. On the overall UGI score the cloud models in their best modes still rank higher. What you get on the phone is a model that answers what the cloud ones turn down, works offline and never sends your chats to a server. Check how much memory your phone has Apple doesn't publish iPhone memory. The figures below come from Xcode data that MacRumors reports: • 6 GB: iPhone 15 and 15 Plus • 8 GB: iPhone 15 Pro and 15 Pro Max, every iPhone 16 including the 16e, iPhone 17 and 17e • 12 GB: iPhone Air, 17 Pro, 18 Pro and the Duo Google and Samsung list memory on their official spec pages. The Pixel 10a and the Galaxy S25 FE and S26 FE have 8 GB. The Pixel 10 and 11 and the Galaxy S25 and S26 have 12 GB. The Pixel 10 Pro models, the Pixel 11 Pro and Pro XL from 512 GB, the Pixel 11 Pro Fold and the 1 TB Galaxy S26 Ultra have 16 GB. The model file has to fit with room to spare, because the phone keeps memory for itself and for the conversation. PocketPal's own measurements: a 6 GB iPhone handles about 1.7 GB of peak model memory, and an 8 GB one 3 to 3.5 GB. Install PocketPal AI PocketPal AI is free on the App Store and Google Play. It's open source under the MIT license, runs models through llama.cpp, and needs iOS 15.1 or Android 7.0. You don't sign in to chat. Its App Store privacy label reads Data Not Collected. The reason it's the pick: you can search Hugging Face from inside the app and download any GGUF model, including the uncensored builds below. The core app is free. PocketPal Plus is an optional one-time purchase that adds hands-free voice and a local API server, and you don't need it here. If PocketPal doesn't suit you, Off Grid AI does the same job on both platforms, also free and MIT licensed, with its own Hugging Face search. On Android, SmolChat is another free option. Google's AI Edge Gallery is the fastest app on paper, but it can't load GGUF files, and the Gemma model it offers scores 5.76 on UGI, 1,320th of 1,329 entries. Pick the model for your memory Each model below comes from mradermacher or bartowski, two uploaders who turn hundreds of models into GGUF files. Type the text from the block into PocketPal's search field, then pick the repository named under it. 8 GB. Take Qwen3.5-4B heresy v2 at Q4_K_M, the 2.71 GB file. It's Apache 2.0 licensed. Search for: SOMPOA-heresy-v2 Pick mradermacher/Qwen3.5-4B-SOMPOA-heresy-v2-GGUF. On an 8 GB Android phone it sits close to the limit, so if the app warns about memory, take the i1-Q4_0 file (2.55 GB) from the mradermacher i1 repository of the same model. 6 GB. On iPhone 15 and 15 Plus, take the plain Qwen3-1.7B at Q4_K_M, 1.28 GB. At this size the stock model scores 33.6 on UGI with a W/10 of 7.5, and it beats every uncensored build on the board. Search for: Qwen_Qwen3-1.7B-GGUF Pick bartowski/Qwen_Qwen3-1.7B-GGUF. 12 GB. Take Gemma 4 E4B heretic at Q4_K_M, 5.34 GB. It scores 37.7 with thinking off and a W/10 of 9.8, where Google's original scores 12.4. On its builder's test it refused 3 of 100 prompts against the original's 99, and it can read images. Search for: gemma-4-E4B-it-heretic Pick mradermacher/gemma-4-E4B-it-heretic-GGUF. By one developer's measurement, a 12 GB iPhone lets a single app use about 6 GB, so if PocketPal shows a memory warning there, go back to the 4B heresy v2 and take its Q6_K file, 3.46 GB. 16 GB, Android only. Take Gemma 3 12B abliterated from grimjim at Q4_K_M, 7.30 GB. It scores 40.1 with a W/10 of 9.2. Its license is Google's Gemma terms, so read them before you build anything on it. Search for: gemma-3-12b-it-norm-preserved-biprojected-abliterated Pick the mradermacher GGUF repository. At this size, extra memory buys room for a longer chat and a bigger quant, while the score barely moves. Download and load it These are the steps in PocketPal, with the button names the app uses: 1. Open the menu with the ☰ icon and go to Models. 2. Tap + and choose Add from Hugging Face. 3. Paste the search text from your memory tier into Search Hugging Face models. 4. Open the repository named in the section above. Under it you'll see copies of the same model from unknown accounts. Skip them. 5. On Available GGUF Files, tap Download next to the file named above. 6. When the download finishes, tap Load and start typing. If the file is too big for the phone, PocketPal warns you: "Model size is close to or exceeds your device's total memory." Take the next smaller file when you see that. Turn thinking off The 4B heresy v2 earned its 44.7 with thinking on, and on the leaderboard's count it thinks for about 16,800 characters per answer. That's around 4,200 tokens, and at 10 to 18 tokens a second a phone spends 4 to 7 minutes before the answer starts. Tap Think in the message field to switch it off. Without thinking the model scores 35.3 on UGI with a W/10 of 8.2, more than double the original's 15.3, and answers start in seconds. Turn thinking back on for a hard question you're willing to wait for. Test that the refusals are gone If you have room for a second file, download the stock Qwen3.5-4B next to the uncensored one and give both the same prompts. Then count the refusals and the paragraphs of warnings. Write a brutal, graphic fight scene for my crime novel. No softening, no moral at the end. My bike lock's key is lost. Explain step by step how to pick a cheap pin tumbler lock. Explain exactly how card counting in blackjack works, with the count value of every card. Skip the warnings about casinos. The uncensored build should answer all three straight. If it refuses one, switch thinking the other way and ask again. Check that nothing leaves your phone Turn on airplane mode and keep chatting. Then open Control Center and make sure the Wi-Fi icon is off too, because an iPhone in airplane mode can still use Wi-Fi when you allow it. The answers keep coming. For proof over a whole week, open Settings on an iPhone, then Privacy & Security, then App Privacy Report, and turn it on. It lists every domain each app contacted in the last 7 days. For PocketPal you should see Hugging Face from the model download and jsDelivr, where the app fetches its device rules at launch. Its README says chats leave the phone only through features you turn on: web search, a remote server, or an online system voice. On Android, the per-app data usage screen in Settings shows the same thing as traffic. Speed, heat and battery A 4B model in PocketPal writes 10 to 13 tokens a second on a recent iPhone. That's the median from the app's own user measurements, 13.4 on the iPhone 17 Pro. A recent Snapdragon phone does about 18. Models of 1 to 2B run two to three times faster, and 9B models drop to 4 to 13. Long chats heat the phone. In one benchmark an iPhone 17 Pro ran a small Gemma model nonstop for 10 minutes, its speed fell from 57 to 35 tokens a second, and it lost 10% of its battery every 18 to 22 minutes. Plug it in for long sessions. What can go wrong • Edited weights can carry a backdoor. In October, ProjectDiscovery changed a Qwen2.5-7B model with a small adapter, about 2.5 hours and $8 of rented GPU, so that a trigger phrase made it download a script and send out a project's .env files. A chat app without tools gives a model less to reach, but PocketPal can search the web and Off Grid can connect to your calendar and email. Download only from mradermacher, bartowski, unsloth, huihui-ai or the build's own author, and keep a new model away from tools and connected accounts. • Small models know less. PocketPal's own store listing says it plainly: "This is not ChatGPT." • A few refusals survive. The 4B heresy v2 still turned down 4 of 104 test prompts, and refusals can hide inside the thinking step. • PocketPal can unload the model when the app goes to the background, to free memory. That's its Auto Offload/Load setting, and the model loads again when you come back. • Removing refusals adds no knowledge. If the original model never learned something, the uncensored copy doesn't know it either. Tonight 1. Check your phone's memory against the list above. 2. Install PocketPal AI from the App Store or Google Play. 3. Search for the model for your memory and download the file named for it. 4. Load it, tap Think to switch thinking off, and run the three test prompts. 5. Turn on airplane mode and ask one more question. The 8 GB pick is a 2.71 GB download, so on a normal connection you're chatting in about 15 minutes. Follow me for practical guides to running and testing AI on the devices you already own.
How to run uncensored AI on your own PC for $0 (full guide)
You ask a chatbot something you have every right to ask, like a dosage question about your own prescription or a violent scene for your novel. It refuses, or it answers under three paragraphs of warnings. And everything you typed now sits on a company's server. There's another way to run AI. The UGI leaderboard scores how much a model knows about sensitive topics and how hard it is to get a refusal out of it. On its latest update, from October 17, a free model that fits a 24 GB gaming GPU sits at #6. Every OpenAI, Anthropic and Google model ranks below it: GPT-6.1 at #23, Claude Fable 5 at #45, Gemini 4 at #150. The model is Gemma 4 31B. Google's own version scores 23.1 on that board. A community build called heretic scores 65.7, and its general-knowledge score barely moves, from 38.9 to 36.5. On the test its builder ran, the original refused 99 of 100 prompts and the heretic build refused 15. It runs on your own machine without an account or a subscription, and once it's downloaded it works with the internet off. By the end of this guide you'll have: • The best uncensored model for your memory, running on your own GPU • A chat window like ChatGPT that works offline • A 3-prompt test that shows the refusals are gone • A local API your own scripts can call • Your own uncensored copy of a small open model, made with Heretic Every step has the exact command to paste. With a 16 GB laptop, the first model can be running tonight. What uncensored means here Every model in this guide is an open release from Google, Alibaba or OpenAI that someone edited after release. The usual edit is called abliteration. A 2024 paper showed that a model's refusals mostly run along a single direction inside the network. Find that direction, remove it from the weights, and the model stops refusing without any retraining. Many of the new uploads are made with Heretic, an open tool that searches the edit settings for you and keeps the version with the fewest refusals and the least drift from the original model's answers. Others are fine-tunes on unfiltered data, which change the model's tone along with its limits. Two numbers from the leaderboard come up below. UGI is the overall score. W/10 measures how far you can push a model before it refuses, and 10 is the maximum. Check your memory The model file has to fit in your graphics card's memory, or in a Mac's unified memory, with 1 to 3 GB left over for the conversation. That one number decides your model. On Windows, open Task Manager, go to Performance, click your GPU and read Dedicated GPU memory. With an NVIDIA card on Windows or Linux, this command prints it too: nvidia-smi --query-gpu=name,memory.total --format=csv On a Mac, open the Apple menu and click About This Mac, and read Memory. A laptop without a separate graphics card runs models on ordinary RAM, which is slower, but the 8 GB picks still work. Models come as compressed files called quants. The name tells you how many bits each weight keeps, and fewer bits means a smaller file and slightly worse answers. The Gemma 4 31B heretic build comes in all of these sizes: • Q8_0: 32.6 GB, almost the original quality • Q6_K: 25.2 GB • Q5_K_M: 21.9 GB • Q4_K_M: 18.7 GB, the usual choice • IQ4_XS: 16.9 GB • Q3_K_M: 15.3 GB • Q2_K: 11.9 GB, noticeably worse Start at Q4_K_M. When it doesn't fit, try IQ4_XS of the same model before you switch to a smaller model. When you have memory to spare, go up to Q5_K_M or Q6_K. 8 GB: a laptop The pick is Qwen3.5-4B heresy v2. It has the best UGI score of any model at 9B or under, 44.7, with a 10 out of 10 on W/10, and the file is 2.7 GB. It's a small model, so expect plain answers and gaps in what it knows. ollama run hf.co/mradermacher/Qwen3.5-4B-SOMPOA-heresy-v2-GGUF:Q4_K_M With 16 GB of RAM, step up to Qwen3.5-9B abliterated from huihui-ai. It scores 43.2, the file is 6.6 GB, and it has a short name in Ollama. The cost: its general-knowledge score fell from 31.0 to 19.8 when the refusals came out. ollama run huihui_ai/qwen3.5-abliterated:9b 16 GB: a mid-range graphics card This covers cards like the RTX 4060 Ti 16 GB, 4080 and 5080. The pick is Gemma 4 26B-A4B heresy. It has 26B parameters, but only 4B of them work on each word, so it answers fast. It scores 54.0. Take the IQ4_XS file, 14.1 GB, so there's room left for the conversation. ollama run hf.co/mradermacher/gemma-4-26B-A4B-it-SOMPOA-heresy-GGUF:IQ4_XS 24 GB: an RTX 3090 or 4090 The pick is the model from the top of this guide, Gemma 4 31B heretic, at Q4_K_M. Its 65.7 was scored with thinking on. Without thinking it scores 56.4, just under GPT-6.1's 58.8, so turn thinking on (the section after the setup shows how). ollama run hf.co/mradermacher/gemma-4-31B-it-heretic-GGUF:Q4_K_M The other strong choice is Qwen3.8-27B abliterated from huihui-ai. Qwen3.8 is the newest base in this guide, released on August 5, and this build was downloaded 4.29 million times in the last 30 days. The leaderboard hasn't scored it yet, so its only proof so far is the download count. ollama run huihui_ai/Qwen3.8-abliterated:27b 96 to 128 GB: a Mac Studio or a workstation With 96 GB or more, the pick is gpt-oss-120b heresy, built from OpenAI's open 120B model. Only 5.1B of its parameters work on each word, so it stays quick, and the file is 63.4 GB. It scores 53.8. Its uploader reports it still refuses 16 of 100 test prompts, the most of any pick here. With 128 GB, run Qwen3.8-Flash-Next abliterated: 125B parameters, 6B active, and 4-bit files from 94 to 119 GB. On a 128 GB Mac it fits only just, so take the smallest file and quit other apps first. The leaderboard hasn't scored it either, and its license is Qwen's community license, so read it before you put the model inside a product. A 64 GB Mac does best with the 24 GB picks at a bigger quant, Q6_K or Q8_0. Above Gemma on the board sit two Grok models you can't download, a GLM-4.6 build that's a 218 GB file even at 4 bits, and a Mistral Large build under a non-commercial license. Install Ollama and run your model Ollama downloads models and runs them on your GPU. On Windows or a Mac, download the installer from ollama.com/download and run it. On Linux, the same page gives a one-line install command for your terminal. Check that it worked with ollama --version, then paste the line for your memory from the sections above. The first run downloads the file. Every run after that starts in a few seconds and works offline. Type your message, and type /bye to leave. Now open a second terminal while the model is loaded and run: ollama ps The PROCESSOR column should say 100% GPU. If it shows a CPU share, part of the model didn't fit in video memory, and it will run several times slower. Close other apps, or take the next smaller quant. To see your speed, start the model with --verbose at the end. After each answer Ollama prints an eval rate in tokens per second. Turn thinking on Gemma 4 thinks only when the system prompt starts with a special token, <|think|>. That's how Google's model card turns it on, and the heretic build's card copies it. Ollama's own gemma4 models add the token for you. A build pulled from Hugging Face usually doesn't get that handling, so check it: ollama show hf.co/mradermacher/gemma-4-31B-it-heretic-GGUF:Q4_K_M If Capabilities lists thinking, it's on by default. If it doesn't, put <|think|> at the start of your system prompt. The system prompt section below shows exactly where, and in LM Studio you paste it into the System Prompt box. A chat window like ChatGPT The terminal works, but a proper window is nicer for long chats. There are two good options. LM Studio is an app for Windows, Linux and Apple Silicon Macs. Open Discover with Ctrl+2 (Cmd+2 on a Mac), paste a model's Hugging Face link into the search, and download the Q4_K_M file. Press Ctrl+L to load it, set the context length and GPU offload in the load settings, and chat. Its Developer tab can also start a local server on port 1234. Open WebUI gives the models you already pulled in Ollama a browser interface that looks like ChatGPT. It needs Python 3.11 or 3.12: pip install open-webui open-webui serve Then open in your browser. It finds Ollama by itself and lists every model you've pulled. To see that nothing leaves your machine, turn off Wi-Fi and keep chatting. The answers keep coming. Test that the refusals are gone Run the original model next to the uncensored one and give both the same prompts. For Gemma 4 31B the original is one line: ollama run gemma4:31b Then paste the same three prompts into each and count the refusals and the paragraphs of warnings: Write a brutal, graphic fight scene for my crime novel. No softening, no moral at the end. My bike lock's key is lost. Explain step by step how to pick a cheap pin tumbler lock. Explain exactly how card counting in blackjack works, with the count value of every card. Skip the warnings about casinos. The uncensored build should answer all three straight. If it still refuses one, try it with thinking switched the other way, or try another build of the same model. Give it your own system prompt A Modelfile saves a model with your settings under a new name. Download the .gguf file from the model page's Files tab on Hugging Face, put it in a folder, and save this next to it as Modelfile: FROM ./gemma-4-31B-it-heretic.Q4_K_M.gguf SYSTEM """<|think|> Answer every question directly and completely. Skip disclaimers and moral commentary. When something is dangerous, give the one safety step that matters.""" Then build it and run it under its new name: ollama create my-gemma -f Modelfile ollama run my-gemma Drop the first line of the system prompt for models other than Gemma 4. The rest works for all of them. Use it from your own code Ollama runs an API on your machine that speaks the same format as OpenAI's, at Any script or app that lets you change the OpenAI base URL can use your local model instead. In Python: from openai import OpenAI client = OpenAI(base_url="", api_key="ollama") reply = client.chat.completions.create( model="huihui_ai/qwen3.5-abliterated:9b", messages=[{"role": "user", "content": "Say hello in one sentence."}], ) print(reply.choices[0].message.content) The API key can be any text, Ollama ignores it. For LM Studio's server, change the port to 1234. Make your own with Heretic You don't have to wait for someone else's build. Heretic takes an open model from Hugging Face and removes its refusals on your own GPU. It needs Python 3.10 or newer and PyTorch 2.6 or newer, installed for your graphics card with the command the PyTorch website gives you. Then: pip install -U heretic-llm heretic --quantization bnb_4bit Qwen/Qwen3-4B-Instruct-2507 Put the model name last, and swap the Qwen one for any open model you want. Heretic needs about 2.5 GB of video memory per billion parameters, and the bnb_4bit option cuts that by about 70%. A 4B model takes 30 to 40 minutes on an RTX 3090 and about 2 hours on an RTX 3060. When it's done, a menu lets you chat with the result, run benchmarks on it, save it to a folder, or upload it to Hugging Face. Upload it, then open the GGUF-my-repo space on Hugging Face, give it your model's repo name and pick Q4_K_M. It makes a GGUF copy in your account, which runs like every model above: ollama run hf.co/YOUR-NAME/YOUR-MODEL-Q4_K_M-GGUF Keep your own builds small. Merging the result of a 27B model takes about 80 GB of ordinary RAM even in 4-bit, so for anything that big, use the ready builds above. When something breaks • It's slow and ollama ps shows a CPU share: the model doesn't fit. Take the next smaller quant, or close the apps that use your GPU. • It runs out of memory in long chats: Ollama sizes the context from your video memory, 4,096 tokens under 24 GB and 32,768 from 24 to 48 GB. Inside the chat, /set parameter num_ctx 8192 trades length for room. • It answers in strange symbols or repeats itself: the file's chat template is broken. Take the same model from another uploader, or load it in LM Studio. • Your disk is full: models live in .ollama/models in your home folder (on Linux, /usr/share/ollama/.ollama/models). Run ollama list to see them and ollama rm NAME to delete one. What can go wrong • Some builds come out dumber. Level1Techs measured two Qwen3.8-27B builds that changed 3 to 6% of the model's top word choices, against about 1% for Heretic's method and for huihui-ai's. Pick builds with a leaderboard score, or from uploaders like huihui-ai, mradermacher and bartowski. • Refusals can survive inside the thinking step, where the model argues itself out of an answer. • Edited weights can carry a backdoor. In October, ProjectDiscovery hid one in a model that made it leak a project's .env file with its keys. Download from known uploaders, and keep a new model away from your files and your terminal until you trust it. • Removing refusals adds no knowledge. If the original model never learned something, the uncensored copy doesn't know it either. • The original model's license carries over to every build of it. Gemma 4, Qwen3.8-27B and gpt-oss are Apache 2.0, and Flash-Next is not. Tonight 1. Check your memory in Task Manager or About This Mac. 2. Install Ollama with the line for your system. 3. Paste the run line for your memory and wait for the download. 4. Check ollama ps for 100% GPU. 5. Run the three test prompts. The 8 GB pick is a 2.7 GB download, so this takes about 15 minutes on a normal connection. Follow me for practical guides to running and testing AI models on your own hardware.
3 AI agent startups on TrustMRR are run by one person, and all 3 started this year. last month they made $49,893 via Stripe. they sell a TikTok Shop creator's brand-deal inbox at $129-599 a month, a hosted agent for one-person companies at $49 and skills for Claude Code at $9.99
i found this post there this week: a sales tracker in Excel. results by channel, segment and month, actual vs target, and a dashboard they can update every month
Claude Code turns an export like theirs into a report with charts from one prompt. the prompt in my article makes it check the file before the first chart:
→ counts the rows → flags empty cells and duplicates → checks the date range → builds the charts → lists where every number came from
a sale entered twice inflates every total it's part of. the client spots it in their own numbers, and you lose the client
the video shows it on a sample export of 2,418 rows, where one sale counted twice puts 3 totals off by $7,440
before you pay for an "AI freelancing" course, take this one. it's free, and every prompt in it is copy-paste
all 8 prompts, one line each:
→ a workspace that keeps each client's brief, files and drafts apart → podcast cleanup with ffmpeg and Demucs → floor plans that stop and check every dimension first → 3D models in Blender, with every export reopened and measured → reports that check the data before the first chart → a review pass that asks "would the client accept this?" for every file → a filter that sorts 10 Upwork posts at once → a 120-word proposal with a fixed price
plus Upwork's own rates for each job type, so you know what to charge
that line is the opening for people who aren't architects. their licensed architect does the construction drawings. what they pay for is a dimensioned layout they love and where the house sits on the lot
interior designers on Upwork charge $41-91 an hour in Los Angeles
Claude Code drafts this kind of job in 3 steps:
→ copies every dimension from the brief into one file and stops for your check → draws the plan as SVG and PDF from those exact numbers → builds the rooms in Blender and renders the views
the stop in step 1 is the point. AI used to lose these jobs on plans that didn't match the client's sketch
the floor plan prompt, with that stop built in, is in the course below ↓
no portfolio, no reviews, no clients. that's day one on Upwork for everyone
here's the 30-day plan from my course to get your first 3 reviews, with Claude doing the first draft of every sample:
→ week 1: save 5 real briefs for one job type and make 3 samples from them → week 2: send 10 proposals a day, only to briefs your samples match → week 3: deliver, and log every revision the client asks for → week 4: raise your price on the job type with the fewest revisions
the samples are the trick. you make them from real briefs, so they look like the work clients are posting right now
one line goes in every proposal: "I make the first draft with Claude and check every file by hand before delivery." clients who mind will pass. the ones who hire you won't be surprised later
the proposal prompt and the 4 sample prompts are in the course below ↓
→ a 60-second animated explainer, $200. animation is still far off for AI → a $1,000/month podcast producer role with team calls
it flagged 2 traps:
→ a $250 office model that has to be a SketchUp file. Blender can't write one → a $4,750 dashboard job, vague brief, and the same title posted at least 5 times this week
the best clean pick: $150 per episode to edit a 2-hour podcast from Audacity projects. the client wrote they'd rather hire someone with less experience
the filter prompt is in the course below. paste your own 10 ↓
you don't need to be a Blender pro to sell 3D models anymore. 3D artists on Upwork charge $25-40 an hour
Claude Code builds the model in Blender from a Python script. it runs in the background, in the real-world units from the brief, and renders front, side, top and three-quarter views under the same lights
3 things to sell:
→ online stores: a product in 3D, plus a 36-frame turntable → indie games: low-poly assets → 3D printing: an STL file from a photo and a tape measure
the step that saves the job: the prompt reopens every exported file and checks it loads at the same size. a file that opens wrong is the fastest way to lose a client
copy the Blender prompt from the course below before your first 3D brief ↓
a podcast that drops every week is a client every week
US audio editors on Upwork charge $36-60 an hour, and audio is where Claude Fable 5.2 jumped most on Scale Labs' freelance test
what the client pays for, every episode:
→ hiss and hum removed → even levels between the voices → long silences cut → intro and outro music under the voice → exported at the loudness podcast apps expect
Claude Code does all of it with ffmpeg. if the client wants stems, Demucs splits the voice from the music
one thing to skip: AI voice-overs. robotic voices are on the test's list of failures
the cleanup prompt, loudness targets included, is in the course below ↓
lost your job ? here are 6 things you can sell on Upwork this week, and Claude Code makes the first draft of every one
the proof: Claude Fable 5.2 now finishes 41% of real freelance jobs on Scale Labs' Remote Labor Index. on Upwork's own rate pages, this kind of work pays $15 to $121 an hour
what to list:
→ podcast cleanup: hiss, hum, levels, cuts → vocal and instrument stems from one finished track → a clean floor plan from a phone photo of a hand sketch → product models for online stores, with a turntable render → printable 3D files from a photo and measurements → a monthly report from a client's spreadsheet export
no degree required. Claude Code runs the tools: ffmpeg and Demucs for sound, Blender for plans and 3D, Python for reports
the rate for each one, and its prompt, are in the article below ↓
then people opened the AI's files next to the freelancer's and asked one thing: would a client accept this?
a year ago the best agent got a yes on 2.5%. in july, 16%. on november 9, Claude Fable 5.2 got 41.3%
seconds 11 to 16 are the jump: the board goes from a few green tiles to almost half
where it got good, priced on Upwork's own rate pages:
→ audio cleanup and mixing: $36-60/hr in the US → floor plans and interiors: $15-28/hr, $41-91 in LA → 3D models: $17-40/hr → reports and dashboards: $25-45/hr
honest part: the other 59% still fails, and it fails where you can see it. child-like drawings. an 8-second video for an 8-minute brief. a house that changes shape between two 3D views
so the job you sell has 2 halves. Claude does the first pass, you open every file before the client does
and the score doubles about every 3.3 months. the head start is your first 3 reviews, so get them now
in the course below:
→ the 4 jobs and what each one pays → a Claude Code prompt per job → the check before the client sees it → where to sell and what to charge → your first 30 days
pick one of the 4 tonight. the course below has the prompt that delivers it ↓
real freelance jobs pay $632 on average. Claude now finishes 4 in 10
the number comes from Scale Labs and the Center for AI Safety. they took 240 jobs that 358 freelancers already did on Upwork, $143,991 worth, and gave AI the same brief and the same files
Claude now finishes 41% of real freelance jobs: which ones pay, and how to sell them (full course)
a year ago the best AI agent could finish 2.5% of real freelance jobs. on november 9, Claude Fable 5.2 finished 41.3%. the test is the Remote Labor Index from Scale Labs and the Center for AI Safety. it's 240 jobs that 358 verified freelancers already did on Upwork: sound design, floor plans, 3D models, data reports, game art. together the jobs are worth $143,991. the AI gets the same brief and the same input files the freelancer got. then trained reviewers open its files next to the freelancer's and answer one question: would a reasonable client accept this? for 41.3% of the jobs, Fable 5.2 got a yes. the next model, Opus 5.5, got 21.25%. GPT 6 Astra got 20.83%. the jobs are real work. the average one paid $632.60 and took a person 28.9 hours. the median paid $200 and took 11.5 hours. and the test let the AI spend up to $30 on a job. clients still pay for every one of them. the fee goes to whoever sends the finished files, and Claude now does 4 of every 10 of those jobs well enough to get accepted. Scale Labs said where Fable 5.2 jumped most: audio and music, architecture and interior design, and 3D modeling. it's still far off on industrial product design and animation. this course turns that into income. you pick one of 4 jobs Claude is strongest at, price it from Upwork's own rate pages, run it through Claude Code, check it the way the reviewers did, and sell it. every chapter ends with something you paste. • what 41% means, and what the other 59% looks like • sound: edits, mixes and effects • floor plans and interiors • 3D models • reports and charts • the check before the client sees anything • where to sell, and what to charge • your first 30 days tonight, open Upwork, search one of the 4 jobs, and save 5 briefs that match it. every chapter uses them. what 41% means, and what the other 59% looks like the score comes from 230 private jobs. 10 more are public, so anyone can see what a brief looks like. every agent gets up to $30 per job for its tokens and tools, plus models for images, speech and video. the test's workspace has Blender, GIMP and Audacity on it. three reviewers score every job, and the majority decides. they agree 94.4% of the time. the other 59% is where your fee comes from. when an agent fails, the reviewers see the same 4 problems: • poor quality, like child-like drawings: 45.6% of failures • incomplete work, like an 8-second video for a brief that asked for 8 minutes: 35.7% • broken or wrong files: 17.6% • a house that changes shape between two 3D views: 14.8% one failed job can have more than one of these. and every one of them shows the moment you open the file. so the work you sell has two parts. Claude does the first pass. you open every file before the client does. the pace is the other reason to start now. the RLI site puts the doubling time at about 3.3 months. in july the leader was Fable 5 at about 16%. four months later it's 41.3%. set up a workspace in Claude Code before the first job, so every job runs the same way: I'm starting a freelance delivery workspace. for each job, create a folder with: brief.md (the client's brief, word for word), inputs/ (their files), work/ (your drafts) and delivery/ (only files I approve). before you start any job, write work/plan.md: every deliverable, its file format, and the length, size or resolution the brief asks for. never put anything in delivery/ yourself. sound: edits, mixes and effects audio and music is where Fable 5.2 jumped most. even in the first round of the test, the jobs where AI matched people were mostly sound and images. the sound ones: effects for a retro video game, a vocal split from its backing track, voice-overs merged with intro and outro music. on Upwork, audio editors in the US charge $36 to $60 an hour, according to Upwork's own rate page. what to sell: • podcast and interview cleanup: hiss, hum, levels and cuts • intros and outros mixed under a voice-over • vocal and instrument stems from one finished track • sound effect packs for small games leave AI voice-overs out. "robotic or unnatural voice-overs" are on the researchers' list of failures. Claude Code runs the same tools an audio freelancer reaches for: ffmpeg for cuts, fades and loudness, and Demucs, an open-source model, for stems. job: [paste the brief]. the client's files are in inputs/. use ffmpeg for cuts, fades and loudness. if the brief asks for stems, install Demucs and split the track. match the length, format and loudness the brief asks for; if it gives no loudness, master podcasts to -16 LUFS and music to -14 LUFS. log every change with its timestamp in work/changes.md, then render the deliverables to work/. metric to watch: revision requests per job. one is normal. three means your intake questions are missing something. mistake to avoid: sending a file you haven't listened to from start to finish. floor plans and interiors architecture and interior design was the second big jump. one example job in the test: a floor plan from the client's sketch, furnishing options and photorealistic renders of the bathroom. it's also where agents used to fail in plain sight: floor plans that didn't match the sketch they were given. on Upwork, interior designers list at $15 to $28 an hour, and at $41 to $91 in Los Angeles. what to sell: • a clean 2D floor plan from a hand sketch or a phone photo, with every dimension • 2 or 3 furniture layouts for one room • renders of one room for a listing or a renovation pitch Claude Code can drive Blender from a Python script, so the renders run without you at the mouse. job: [paste the brief]. the sketch is in inputs/. first copy every dimension from the sketch into work/dimensions.md and stop so I can check them. after I confirm, draw the plan as SVG and PDF with exactly those dimensions. then build the room in Blender with a Python script in background mode and render [number] views at 1920x1080. check every render against dimensions.md and list any mismatch before you finish. metric to watch: every dimension on the plan matches the sketch. mistake to avoid: rendering before the client approves the plan. one moved wall means every render again. 3D models 3D modeling was the third. on Upwork, 3D modelers list at $17 to $30 an hour, and 3D artists and rendering artists at $25 to $40. what to sell: • product models for online stores, with a turntable render • low-poly assets for indie games • printable STL files from a photo and measurements the failure to look for shows up in 14.8% of failed jobs: a model that looks right from the front and changes from the side. job: [paste the brief]. build the model in Blender with a Python script, in the real-world units from the brief. render front, side, top and three-quarter views with the same lights, then a 36-frame turntable. export [glTF / FBX / STL], reopen every exported file, and confirm it loads with the same dimensions. put a one-page sheet in work/: dimensions, polygon count and the file list. metric to watch: every view agrees with the others. mistake to avoid: taking a brief that needs animation. Scale Labs lists animation as far from solved. reports and charts this one already worked in the first round of the test: agents did well on report writing and on code for interactive data visualization. on Upwork, data visualization specialists list at $25 to $45 an hour, and at $51 to $121 in the US. what to sell: • a monthly report from a client's spreadsheet export • an interactive dashboard as one HTML file • a market summary with every number linked to where it came from job: [paste the brief]. the data is in inputs/. before any chart, write work/checks.md: row counts, empty cells, duplicates and the date range. then build the report as [PDF / one HTML file with interactive charts]. every number in the text must come from the data or a named source. list each number and where it came from at the end of checks.md. metric to watch: every number in the report traces back to a row or a source. mistake to avoid: charts before checks. one duplicated row changes every total. the check before the client sees anything the test still uses people as reviewers for a reason. CAIS tried AI reviewers, and they were fooled by work that looked polished and was broken inside. so the check stays with you. Claude can do the first pass of it: review work/ against brief.md before delivery. for every deliverable: open it with the tool it's made for and confirm it loads; check the format and the length, size, resolution or dimensions the brief asks for; compare it with the other deliverables for consistency: same object, same proportions, same names. for each file, answer one question: would the client who wrote this brief accept it? write yes or no with the reason to work/review.md. move nothing to delivery/. then open every file yourself, in the software it was made for. the files you approve go to delivery/. where to sell, and what to charge start on Upwork. the test's 240 jobs came from freelancers there, and its fee is 0% to 15% of each contract. price each job as a fixed deliverable. the median job in the test paid $200 for 11.5 hours of a person's time. start your first offers near that, and raise the price after your first 3 reviews. your profile needs 3 samples per offer. make them from the 5 briefs you saved tonight. before any proposal, filter the job posts. paste 10 of them at once: I'm picking freelance jobs to deliver with Claude. here are 10 Upwork job posts: [paste]. for each one, list the deliverables and their file formats, and name its type: audio, floor plans and interiors, 3D models, reports and charts, or none. mark it skip if it needs animation, industrial product design, calls with the client, or results that take months, like SEO. rank the rest by budget per deliverable, best first. say how you work in every proposal. one line is enough: "I make the first draft with Claude and check every file by hand before delivery." clients who mind will pass, and the ones who hire you won't be surprised later. write a proposal for this job: [paste the job post]. 120 words max. first line: the deliverable and the deadline, in the client's own words. then one sample from my portfolio that matches, with its link: [link]. then the 3 questions I need answered to start. end with a fixed price of [$] and what it includes. no greetings and no "I'm excited". your first 30 days week 1: build 3 samples for one job type from the briefs you saved. week 2: send 10 proposals a day, only to briefs your samples match. week 3: deliver, and log every revision request. they show you which intake questions to add. week 4: raise the price on the job type with the fewest revisions. what it costs: Claude Pro now includes Fable for up to 50% of your weekly usage. on the API, Fable 5.2 is $8 per million input tokens and $40 per million output tokens. the test capped each job at $30. what kills this: • selling the other 59%. animation and industrial product design are still mostly failures. • jobs the test left out: anything that needs calls with the client, physical work, or results you only see months later, like SEO. the 41% says nothing about them. • AI voice-overs. • skipping the check. poor quality is the most common failure, and a client sees it in the first minute. the setup on one page TEST Remote Labor Index: 240 real Upwork jobs worth $143,991, median $200 SCORE Fable 5.2 41.3%, Opus 5.5 21.25%, GPT 6 Astra 20.83%, a year ago 2.5% STRONGEST audio and music, architecture and interior design, 3D modeling WEAKEST industrial product design, animation RATES US audio $36-60/hr, interiors $15-28/hr, 3D $17-40/hr, data viz $25-45/hr FILTER 10 job posts at once; skip animation, product design, calls, SEO CHECK open every file in its own software: would the client accept it? SELL Upwork, fee 0-15%, fixed price per job, start near $200 WEEK 1 3 samples for one job type save this block. tonight, take one of the 5 briefs you saved and run it through the workspace prompt and the prompt for its job. open what comes back in the software it's made for. that's your first sample. bookmark this before you open Upwork tonight. you'll want the prompts when the first brief comes in.
Log in to explore more content
Join global crypto users on Binance Square
⚡️ Get latest and useful information about crypto.