walid@portfolio:~/lab$
cd../portfolio
The Lab

Ideas & Experiments

Things I am building, testing, or arguing with. Half of these will not survive contact with reality — that is the point of writing them down while they are still running.

56 logged1 in flightlast updated Oct 2026
01ideaagentsOct 2026

An AI-native social media agent team

One orchestrator, eight departments, eight interrupts and one owner who approves: the full board, how the loop runs and the Claude Code project behind it, as published, with the board as pictures and all eight files the page shows, checked against Claude Code’s docs and each tool maker’s own

A one-screen design for running a company’s social media with AI agents, and the Claude Code project behind it: one orchestrator, eight departments, eight events that interrupt the loop, and one owner who approves. Here it is as published, with the board as pictures and all eight files the page shows. The checks found a design and a starter template, not a running system. As published it connects to no tools and nothing runs on its own; the page shows 8 of the project’s 55 files; its approval rules contradict each other; and three tool claims are wrong: Metricool doesn’t rotate ads, Midjourney has no official API, and Grammarly’s APIs return no corrections.

Claude CodeSubagentsSocial mediaOrchestrator+1read →
02ideamediaOct 2026

Finding the small YouTube channels that are breaking out

A free script, a YouTube API key and Claude Code scan the last ten days of YouTube for small faceless channels whose videos beat their own numbers, then the same tools help you make one. The scanner, reviewed and hosted here, the eleven channels it found, the toolbox, the steps and the 30-day plan, checked against YouTube’s own rules, the channels’ public numbers and the tools’ licences

A 5 KB script, a free YouTube API key and Claude Code scan the last ten days of YouTube for small channels whose new video beat their usual numbers many times over. The guide adds a toolbox, the steps to make a narrated documentary of your own and a 30-day plan. The scanner works, and a reviewed copy is on this page. Much of what the guide builds on it didn’t hold up. The free API now allows 100 searches a day, and the headline “27x” outlier is really 270x. The eleven “faceless” channels include an on-camera host posting four videos a day, unlabelled AI footage and reused documentaries. YouTube’s monetization rules name animal-peril series as unmonetizable, and ElevenLabs’ free voices can’t go on a monetized channel. All of that is marked below.

YouTubeClaude CodeYouTube Data APIFaceless channels+2read →
03ideaagentsOct 2026

Thirteen Claude skills that run an Instagram account

A free, open-source skill pack that writes your Reels, captions, carousels, replies and DMs in your own voice once you give it your voice profile, scores your profile and finds what is working in your niche, then leaves the posting to you. The full setup for the Claude app and Claude Code, every prompt, and what each skill really does, checked against the pack’s own code, Anthropic’s docs and Instagram’s help pages

Thirteen free Claude skills for running an Instagram account. /ig-reel turns one idea into three scored hooks, a script and a timed beat sheet. /ig-viral ranks the Reels in your niche by how far each one beat its own account. /ig-human strips the stock AI phrasing before anything goes out. Nothing is posted unless you post it. The checks kept the workflow and fixed the setup. Pasting the link only installs it in Claude Code, not the Claude app; the manual install command needed a fix; only two skills read your voice profile on their own; and cleaning up a caption doesn’t change whether Instagram labels a video as AI.

Claude skillsClaude CodeInstagramReels+1read →
04ideamediaOct 2026

A free AI video generator that runs offline

LTX-2 inside ComfyUI on your own NVIDIA graphics card: a prompt or a single image becomes a clip with its own sound, camera moves, effects and any style, with no credits and no watermark. The full ten-step setup and every prompt, checked against Lightricks’ model card and licence, ComfyUI’s own template files, and Instagram’s video specs

Install ComfyUI’s free desktop app, load the LTX-2 template, and a prompt or a single photo becomes a clip with its own sound: engine rumble, rain, a piano note, made in the same pass as the picture. No subscription, no credits, no watermark, and once the models are downloaded, no internet. The ten steps and every prompt survived the checks. The numbers around them didn’t. LTX-2 is open weights rather than open source, free below $10 million a year in revenue. It wants far more graphics memory than “8 GB minimum”, and none of the per-clip times are published anywhere. The model files take 45 GB, not 40. And half the ComfyUI menus the guide names have moved. All of that is fixed below.

LTX-2ComfyUIOpen weightsVideo+2read →
05ideatoolingOct 2026

A bunker of free AI models on your own drive

Six free models and apps, plus Ollama to run them, with the exact commands: Qwen3.5, gpt-oss, GLM-5.3, Jan, AnythingLLM, and Ollama inside Claude’s desktop app. All but GLM-5.3 run on your own computer; its easy route is Ollama’s cloud. A memory guide, a backup kit, and what the headlines behind it actually were. Checked against the makers’ pages, the licences and Ollama’s own docs

A model on your own drive can’t be taken back. This is the short list of free ones worth having, with the exact command for each: install Ollama, pick the model your memory can hold, add Jan for a ChatGPT-style window and AnythingLLM to ask questions of your own files, and, on a Mac, run them inside Claude’s desktop app. The checks kept every command and fixed the story around them. The model “they just pulled” was never pulled: Hugging Face disabled one unofficial copy of GLM-5.3 with its safety refusals removed, and Z.ai’s own release is still up. Qwen3.5’s 35B doesn’t run on a phone the normal way, GLM-5.3’s cloud route isn’t local and may cost money, and the “10-minute” backup kit includes a 330 GB download.

OllamaLocal AIOpen weightsQwen3.5+2read →
06ideamediaOct 2026

Motion graphics as code, timed to a voiceover

A starter kit for explainer animation with no After Effects: every scene a TypeScript function of time, word-level timings from forced alignment, a headless-Chrome renderer and a scripted sound mix. The kit’s full workflow guide, its demo prompt, and the 92-second video it renders, with sound. Built on the open-source pdoom-video engine, and checked against its code

An explainer video in which nothing is keyframed by hand. Each scene, called a plate, is a TypeScript file that draws every frame from the clock; word-level timings of a voiceover tell each plate when to move, so the picture stays locked to the voice; a script drives headless Chrome to render the frames and ffmpeg turns them into an MP4. Because it is all code, you change the animation by asking Claude to change the code. This is the kit’s whole workflow guide and the 92-second video it renders, sound included. The kit’s own files weren’t part of what was shared, but its engine is public, so the preview and render half works from the open-source pdoom-video today. The checks found the engine matches the guide closely, and fixed two example commands that point at the wrong moment.

Motion graphicsClaude Codethree.jsTypeScript+2read →
07ideamediaOct 2026

An AI production studio for full YouTube videos

OpenMontage from an empty laptop: a paid Claude plan, four free helper programs, five setup steps on Windows or Mac, seven copy-paste prompts for full videos, Shorts, ads and documentaries, the pro moves, the mistakes that cost money, and a first-week plan. Checked against the project’s own code, Anthropic’s docs, the providers’ pricing and YouTube’s policies, with the steps that no longer work fixed

OpenMontage turns an AI coding assistant into a whole video crew: it researches, writes the script, plans the scenes, finds real footage or generates visuals, narrates, adds music and captions, edits and renders an MP4, and stops for your OK at every creative decision. This is the beginner’s route in, from a laptop that has never opened a terminal to a first finished video, with every command and all seven prompts. The checks turned up a handful of steps that no longer work as written. Claude Code now starts in auto mode, so it doesn’t ask before each command unless you switch it; on a Mac, Homebrew’s Python 3.12 isn’t the python3 the setup finds; Piper installs without a voice; new Pexels keys are paused; the budget limits are rules the agent follows, not locks in the code; and the free ElevenLabs tier can’t be used in anything you monetize. Each fix is marked where it applies.

OpenMontageClaude CodeYouTubeVideo+2read →
08ideatradingOct 2026

Jev picks the coin, the Grok Bots trade it

A memecoin desk where every judgement is a typed number. A collector pulls fresh launches on Solana, BSC and Robinhood Chain, TypeSafe’s Jev answers yes-or-no, pick-one and score questions about each, and the Grok Bot seats size, fill and exit the one it picks. Every file and prompt from the guide, the overnight $100 run from its video, and what the checks found

A memecoin trading desk rebuilt so nothing on it reasons in prose. A collector pulls fresh launches on Solana, BSC and Robinhood Chain and kills most of them with free arithmetic; TypeSafe’s Jev answers typed questions about the survivors (is the float concentrated, is this X account really the project’s, has the move already happened) and then picks at most one; and the Grok Bot seats size it, fill it on the FOMO app and exit on a single rule. Every file and prompt from the guide is here, run end to end against the real SDK. As published the code can’t trade a single Solana token, so the bugs that stop it are fixed in place and marked. The companion video of a $100 overnight run to $2,197.67 is here too: its coin’s chart checks out, but the run is a replay of candle closes that bets 54 to 81% of the bank on each trade, where the guide’s own rules allow 6%.

JevGrok BotMemecoinsTypeSafe+2read →
09ideatoolingOct 2026

Run an uncensored AI model on your own computer

A free, offline setup for an abliterated open-weights model with Ollama: install it, pick the size your graphics card can really hold, fix the context-length trap nobody mentions, and keep the port off the internet. Plus what “uncensored” actually means, and where it falls down. Checked against Ollama’s docs, the model’s page and the original research

A model on your own machine has no account attached, no usage meter, no network connection after the download and, if you pick an abliterated one, no list of things it won’t discuss. This is the whole setup with Ollama and an abliterated Qwen 3.5: install the runtime, make sure its server is actually running, choose the largest model your graphics memory can hold, and raise the context window Ollama quietly keeps small. It is honest about the trade. A hosted frontier model is smarter than anything a laptop runs, and removing refusals adds willingness, not capability. The reasons to do it are privacy, offline use, and questions a hosted filter blocks for no good reason.

OllamaLocal AIOpen weightsQwen 3.5+2read →
10ideaagentsOct 2026

Inside an AI shell assistant: the Angela CLI blueprint

How a Gemini-powered terminal assistant is designed to fit together: shell hooks, a context engine, a request pipeline, a five-level risk classifier, adaptive confirmation and a rollback log. The design is worth borrowing; the project behind it is an unfinished prototype, and the checks say exactly where. Checked against the project’s own code

Angela is an AI assistant that lives in your shell. You type what you want in plain English; it works out the command, scores how sure it is, rates how risky it is, previews it, asks before anything destructive, runs it, and can roll it back. That loop is a clean blueprint for any AI that acts on a real machine, built from textbook parts: shell hooks that watch what you do, a context engine, a dispatcher, a five-level risk ladder, confirmation in proportion to risk, and a transaction log for undo. This is the blueprint, with each part checked against the code. The honest finding: the architecture is worth borrowing, the project is not something to rely on. It is an alpha prototype, untouched since May 2025, pinned to a Gemini model Google has since shut down, and several of the pieces the design describes are missing or broken.

AI shellArchitectureGeminiPython+2read →
11ideamediaOct 2026

A library of 36 fashion-model poses

Thirty-six ready-to-paste pose prompts for an AI image model — each a full-body studio or editorial shot driven from a reference photo, with the picture it produces. Part one of a three-library system; camera angles and movements are the other two

A grid of 36 fashion poses, each card showing the picture, the exact prompt that makes it, and a copy button. Every prompt keeps the same shape — What (the subject, read from your reference image), Pose (the one thing that changes), Location (studio or editorial lighting) and Additional (quality tags) — so you can swap the reference and keep the pose, or swap the pose and keep the look. Filter by the three categories: Standard catalogue poses, Editorial poses, and a Test set of livelier shots. The companion libraries add camera angles and camera movements.

AI imageFashionPromptsPose library+1read →
12ideamediaOct 2026

A library of 35 camera angles

Thirty-five camera-angle prompts that re-photograph the same subject from a new height, framing, orientation, position, lens or aperture — each with the shot it produces. Part two of the fashion-prompt system

Thirty-five camera angles, each card a prompt that keeps the subject, outfit, lighting and location identical and changes only the camera, with the picture and a copy button. Filter by what moves: height (eye level, high, low…), framing (close, wide…), orientation, the camera’s position, the lens, and the aperture. Every prompt opens with the same “re-photograph, change only the camera” instruction, so you can run the exact same model through any angle. Pairs with the pose and movement libraries.

AI imageCamera anglesPromptsFashion+1read →
13ideamediaOct 2026

A library of 15 camera movements

Fifteen camera-movement prompts for AI video — crash zooms, dolly moves, orbits, crane and handheld shots — each with its five-part recipe and a still from the demo clip. Part three of the fashion-prompt system

Fifteen camera movements for AI video, each card a five-part recipe — the Move, the Camera action, the Pace, what to Lock (keep stable), and how to Resolve (where the shot ends) — with a still from the demo clip and a copy button. Filter by family: zoom, dolly, pan/tilt, orbit, crane/drone and handheld/POV. Each has a duration; the live library plays every clip. Completes the pose and angle libraries.

AI videoCamera movementPromptsFashion+1read →
14ideatoolingOct 2026

Let Claude drive DaVinci Resolve

A safe, step-by-step setup for the community DaVinci Resolve MCP server, so Claude Code or Codex can inspect a project and run real timeline, media, Fusion, grading and render work. Every agent prompt from the guide, plus the one habit that keeps it from wrecking your edit. Checked against the project’s own package

Connect an AI agent to DaVinci Resolve and it can inspect your project and do real work: cuts and trims, markers, timeline versions, media bins, Fusion graphics, supported grades and render jobs, all from plain-English prompts. This is the whole setup — install, enable scripting, the one installer command, the diagnostic, and wiring up Claude Code or Codex — followed by every test prompt from the guide, ordered from a read-only inspection up to a silence cleanup and a render. It is the community davinci-resolve-mcp project, not an official Blackmagic tool, and it drives Resolve through Resolve’s own scripting API, which Blackmagic gates to the Studio edition. The discipline that makes it safe is one habit: inspect, then act, then verify, always on a duplicate.

DaVinci ResolveMCPClaude CodeCodex+1read →
15ideatoolingOct 2026

A toolkit, a dark-web search engine, and an email lookup

Three reconnaissance tools security people actually use: Z4nzu’s hackingtool, the Ahmia search engine for Tor onion services, and the emailOSINT reverse-email lookup. What each one does, when to reach for it, and the single line that separates research from a crime. Checked against each project’s own pages

Reconnaissance is the first phase of any security assessment: before you test a system you map what is already exposed about it. These three tools each cover a different surface. hackingtool is an all-in-one launcher that installs and runs 215 open-source security tools from one menu, now wrapped in an AI layer. Ahmia is a search engine for the Tor network, run openly since 2014, that indexes onion services and filters abuse material. emailOSINT is a reverse-email lookup that surfaces the accounts and breaches tied to an address. All three are legal to run and useful for defenders, researchers and learners. All three are also only lawful when pointed at systems and people you own, have consent for, or are authorised to test. That line, not the tool, is what matters.

OSINTReconSecurityPentest+1read →
16ideatradingOct 2026

Claude researches, Jev decides, you approve

A simulation-only trading research loop: Claude writes the analysis, TypeSafe’s Jev answers TRADE, WAIT or REJECT, a risk check sizes the plan, and nothing happens until a person says yes. Its script was run against both SDKs, so the crash on Claude Opus 5.5 is fixed and the risk check it leaves out is written

A trading research system that never places a trade. Claude reads only the market data you give it and writes facts, assumptions, gaps, a thesis and risks. Jev, TypeSafe’s decision model, turns that into TRADE, WAIT or REJECT with probabilities and a confidence. A risk check sizes any TRADE inside your own limit, and every decision is saved as a JSON file that waits for a person to approve it. This is the whole guide: setup, script, four prompts, a worked example, safety rules, troubleshooting and a final checklist. Its script was also run against both SDKs. On Claude Opus 5.5 it crashes whenever Claude thinks before answering, the risk limits it sets are never read, and its safety checks vanish under python -O. The fixed script and a working risk check are below.

Claude APIJevTypeSafePaper trading+1read →
17ideamediaOct 2026

The kids’ cartoon channel your PC animates

Claude writes every Short, and Wan2GP, a free studio that runs open video models on your own graphics card, animates it. The full system with its prompts, checked against YouTube’s made-for-kids rules, the model licences and the Partner Program changes coming in February 2027

A faceless 3D cartoon channel for kids on YouTube Shorts. Claude writes each Short and its scene prompts, and Wan2GP animates them on your own graphics card, so there are no credits and no fee per clip. This is the whole system: the niche, the install, the models, a character sheet, a week of scripts, a way to learn from the biggest channels and a weekly schedule. It is also checked against the tools’ own pages, YouTube and the FTC, and several things change. Kids’ videos earn less per view and face extra quality rules. The Flux model the guide suggests and CapCut’s music aren’t licensed for a paid channel. Kling’s free daily credits now go to subscribers only. And from February 2027, a new channel needs twice the views to join the Partner Program.

YouTube ShortsWan2GPKids contentLocal video+1read →
18ideaagentsSep 2026

Seven jobs for a model that only decides

TypeSafe’s Jev never writes a sentence. It picks, scores and answers yes or no about text, for $0.042 per million tokens. Seven ways to use it, each with the exact question, the setup in ManyChat, Clay, Zapier, LangChain or LiteLLM, and what it costs

Jev is a model with one job: deciding things about text. Give it some text and a question and it picks an option, places the text on a scale or answers yes or no, always with probabilities and never with a sentence. That is why it is fast, and why it costs $0.042 per million input tokens with output free. Seven ways to put it to work follow, from finding the buyers in your Instagram comments to stopping an agent before it emails 4,212 contacts. Each comes with the exact question to ask, the setup and the cost, and the starter kit’s code is all on the page. Checked against the sources it cites, most of the numbers hold. Three don’t: the cascade result it headlines is within noise, invoices aren’t Jev’s weakest area, and Clay can’t hold 100,000 leads in one table.

JevClassificationLead scoringGuardrails+1read →
19ideamediaSep 2026

The automated explainer channel

Claude Fable finds and scripts the topics, Higgsfield renders the videos, and a schedule keeps it going five days a week. Checked against YouTube’s current rules, which decide whether a channel like this gets paid at all

A faceless YouTube channel that mostly runs itself. Pick a niche by what it pays and find videos that already won. Claude Fable writes a better script, Higgsfield turns it into a narrated, animated video, and a scheduled task keeps new scripts coming. This is the full system, with both prompts. It is also checked against what YouTube, TikTok, Anthropic and Higgsfield say today, and that changes the money more than the method. YouTube now refuses to pay mass-produced AI channels, or AI “experts” giving finance or health advice. Short clips earn nothing on TikTok, and from February 2027 YouTube Shorts pay only past 10 million views in 90 days.

YouTubeClaude FableHiggsfieldFaceless video+1read →
20ideatradingSep 2026

Inside the five-minute Bitcoin bots

How the most profitable Polymarket bots price, build and size positions in “Bitcoin Up or Down 5m”, checked against the wallets’ own numbers and against a rule change that landed the day this was published

Every five minutes, Polymarket opens a market on whether Bitcoin will finish the window higher than it started. It looks like a coin flip. The bots that dominate it treat it as a pricing, execution and inventory problem instead, and this is the full breakdown of how, with the code. Checked against the wallets’ own records, the strategies hold up but the headline doesn’t. Of six bots that could be checked, three made $100K in the month before the article and one has kept that pace since. The market’s settlement rule also changed the day the article was published.

PolymarketMarket makingArbitrageKelly+1read →
21ideaagentsSep 2026

Five hidden rulebooks, two already public

The system prompts behind Claude, Codex, Gemini, ChatGPT and Grok. Each one has a few hundred words that decide how the model writes, checked line by line against the files and against what the vendors publish themselves

Every chat assistant reads a long set of instructions before your first word. A public repository collects them, and five recent ones from Anthropic, OpenAI, Google and xAI are below. A few hundred words in each shape how that model thinks and writes, and those passages are worth borrowing. They are all here, along with a single system prompt built from all five. Checking them turned up something the “leak” framing hides: the Claude passage is published by Anthropic, and most of the Codex one appears in OpenAI’s own developer guide.

PromptingSystem promptsWritingClaude+1read →
22ideamediaSep 2026

Sell the edit, check the client

A free local video engine your coding agent drives, a price list and outreach scripts for selling what it makes, and the licence clause that decides which clients you can take

Mirage, the company behind the Captions app, released Tesseract: a local video engine your coding agent drives directly, with cuts, titles, motion graphics and sound, and no Premiere, After Effects or credits. The agent plan you already pay for can now do real editing work, and people pay for that work. This is the full playbook for installing it, pricing the edits and landing the first clients. It also covers the part the playbook leaves out: Tesseract is free, but not for client work once the client’s business passes $1M a year.

VideoAgentsFreelancingPricing+1read →
23ideamediaSep 2026

The engine on your own machine

Running an open-weights video model locally — what the hardware actually has to be, and what the licence lets a business do with the output

LTX-2.5 turns a written description into a clip with sound, and Lightricks published the weights — so it runs on hardware you own rather than being rented a clip at a time. Multi-shot with a consistent character is the headline. The parts that decide whether it is usable are less glamorous: a hard VRAM floor, 160GB of disk before the first frame, and a licence with a revenue line and a restriction most summaries leave out.

Open weightsVideoLicensingLoRA+1read →
24ideamediaSep 2026

Six agents, one stage each

A video pipeline from raw idea to edit plan — and the gap between rules written in a prompt and rules a tool can enforce

Six prompts, each owning one stage of making a video, chained so every stage consumes the approved output of the last. The discipline is the good part — one job per agent, a human gate after the first two, and a standing ban on inventing numbers. The part worth completing is that Claude Code has a native form for exactly this, and it can enforce the safety rules the prompts can only request.

Claude CodeSubagentsVideoPipeline+1read →
25ideatradingOct 2026

The bot that pays its own subscription

An always-on agent trading prediction markets on a ten-minute loop — a sound system attached to a headline number that cannot survive its own rules

An agent on its own cloud machine scans prediction markets every ten minutes, reads X for sentiment, prices its own fair value, and bets half-Kelly capped at 6% of the bankroll. The risk framework is genuinely good and the Kelly arithmetic checks out. The $50-to-$5,273-in-48-hours claim wrapped around it does not — not against the position limits the same guide sets.

PolymarketAgentsKellyPrediction markets+1read →
26ideatoolingSep 2026

Seven open releases, two licence traps

What OpenAI, xAI, NVIDIA, Anthropic, Cloudflare, Tencent and Alibaba put on GitHub — and which ones you can actually bill for

Seven things the big labs open-sourced, all of them running on your own hardware rather than someone else’s API. Five are plain Apache 2.0 and you can put them straight into client work. The other two are where the actual question lives: one ships MIT code with separately licensed weights, and one excludes the EU, the UK and South Korea outright.

Open sourceLicensingAgentsVoice+1read →
27ideaagentsSep 2026

Two hundred specialists, install five

A roster of role-defined agents for your coding tool — and why the restraint matters more than the catalogue

Instead of one generic assistant, a library of specialists — each with a defined role, workflow, deliverables and success criteria — that install into your coding tool as agent files. The catalogue is the headline; the discipline is the point. Install three to five that match your actual week, give each a concrete deliverable, and chain their outputs rather than treating them as separate chatbots.

AgentsClaude CodeMITRoster+1read →
28ideaagentsSep 2026

Give an agent $100 and a shutdown switch

A local app where every model call costs the agent money, and hitting zero ends the run

An agent starts with a fictional $100, researches real problems, picks one opportunity, builds the assets and writes a sell plan. Every call it makes is deducted from its balance. At zero it stops. The design move worth studying is not the earning — it is that revenue can only be entered by a human with proof, so the one number that would make the whole thing look successful is the one number the agent physically cannot touch.

AgentsPythonFlaskConstraints+1read →
29ideamediaSep 2026

One key, every video model

Wiring a hosted video agent into your coding agent — it plans the shots, picks the model per shot, and returns a finished MP4

One install and one API key put a hosted video agent behind your coding agent: you describe a video in conversation, it writes the script, picks a model per shot, generates the clips, adds music, narration and subtitles, and hands back a single finished MP4. The skills are free and MIT; the generations cost credits, and nothing paid runs without you approving an estimate first.

VideoAgent SkillsClaude CodeMIT+1read →
30ideaagentsSep 2026

Make one job repeatable

A folder, a ten-section playbook, and four prompts that build it, equip it, run it and grade it

The gap between a good one-off result and a system is that nobody wrote down what good looks like. This is one folder and one file that does: purpose, success criteria, rules, output format — plus the three sections most people skip, which are where the value actually is. Four prompts build the playbook, work out what it is missing, run it, then grade the result against it and say which of five things went wrong.

PromptsProcessVerificationTemplates+1read →
31ideaagentsSep 2026

A software company in one command

Agents with job titles — product manager, architect, engineer — turning one sentence into a PRD, a design and code

Instead of one chatbot doing everything, each agent gets a job title and a deliverable: the product manager writes requirements, the architect writes the system design, the engineer writes the files. You type one sentence and get a folder with all three. It costs real money per run, ships with a $3 cap for a reason, and the honest framing is that nothing here promises working software on the first try.

Multi-agentPythonOpen sourceMIT+1read →
32ideatoolingSep 2026

Rebuild it, or keep paying

A catalogue that rates 1,093 paid tools on whether an agent could replace them — and why 86% of the time it should not

Someone went through 1,093 paid products and judged, one by one, whether a coding agent could rebuild each of them. Only 14% got the top verdict. That number is the entire value: the catalogue is far more useful as a list of what not to build than as a shopping list, because it is the rare place that will tell you no with a straight face.

SubscriptionsClaude CodeSQLiteOpen source+1read →
33ideamediaSep 2026

Unlimited, because it runs on your machine

Setting up a free local app that runs open video, image and audio models on your own GPU

One app, installed with a click, running open video, image and audio models locally. The trade is explicit: you supply the GPU and the patience — several minutes per clip on a modest card — and in exchange there is no subscription and no generation limit, forever. The hardware check comes first, because this is the one setup where the answer might genuinely be “not on your machine”.

VideoLocalGPUOpen models+1read →
34ideatradingSep 2026

Six analysts, one report

A research desk built entirely out of prompts — scanner, chart, news, quant, risk, plan

No code, no broker connection, no execution — six prompted analysts that run in strict order over data you paste, and hand back one report ending in READY FOR HUMAN REVIEW. The design discipline is the interesting part: every analyst is forbidden from inventing a number, and one of the six exists purely to try to kill the idea.

TradingPromptsResearchRisk+1read →
35ideatoolingSep 2026

Seven repos worth the disk space

Memory, a context layer, a real browser, orchestration patterns, security and science skills

Seven open-source projects that bolt onto a coding agent: persistent memory, a unified context store, a browser it actually drives, the orchestration patterns that hold it together, and skill libraries for security and science. Every name and star count re-checked — one is commonly mislabelled as official when it is not, and one name collides with a different repository entirely.

AgentsClaude CodeSkillsOpen source+1read →
36ideaagentsSep 2026

A sales agent that stays up

WhatsApp in, AI reply out — with memory, lead scoring, human handoff and a VPS that keeps running

Not a chatbot script — a small production system with separate jobs for messaging, reasoning, memory, lead state, deployment and optional voice. The separation is the whole point: it makes the thing debuggable, and it means a front-end change cannot take down your customer messaging. Ten steps from a local clone to a worker that survives you closing the laptop.

WhatsAppSupabaseNode.jsVPS+2read →
37ideatoolingSep 2026

Website in, React app out

Standing up the open-source site-to-React builder, and the five things it does not ship with

Point it at a URL, get a modern React app you can then edit by chatting at it. MIT licensed, roughly fifteen minutes to run locally, and the setup below is taken from the project’s own README rather than second-hand. Worth being clear-eyed about two things: upstream calls it an example app, not a product, and what it hands you is a builder with no login, no billing and no usage cap — which is the difference between a demo and something you can charge for.

ReactMITSelf-hostedScraping+2read →
38ideaagentsSep 2026

One command, one local agent

Running DeepSeek’s open-source agent harness on your own machine

An open-source agent harness you run yourself — the agent loop, the tools and the web interface all execute locally, and every piece is swappable as a plugin. One npx command puts a UI on localhost. The install itself is genuinely a one-liner; the parts worth reading twice are the developer-preview warning, the package-name collision, and the fact that free software does not mean free model calls.

AgentsOpen sourceSelf-hostedPlugins+1read →
39ideamediaSep 2026

The agent that makes the video

Setting up an open-source video production pipeline your coding agent directs end to end

You describe a video in one sentence and the agent runs research, script, scene plan, assets, edit and render — pausing for your approval at three gates and telling you the cost before it spends anything. There is a genuine zero-key path: free narration, free archival footage, free composition. One correction to the guide this came from: the project is AGPL-3.0, not MIT, which matters a great deal if you plan to build a business on top of it.

VideoClaude CodeRemotionFFmpeg+2read →
40ideaagentsSep 2026

Nine agents, one voice

A voice-driven trading desk where a master agent routes your spoken command to nine specialists

Most people cloning Jarvis build something general. This one goes narrow and deep into a single domain: you speak, a master agent routes the work to nine specialists — macro, smart money, recon, analyst, technician, quant, risk, execution, journal — and the charts on the wall obey out loud. The architectural idea worth stealing is the gate: analysis cannot execute, and execution cannot authorise itself.

AgentsTradingVoiceClaude Code+2read →
41ideaagentsSep 2026

The agentic backend

Every agent demo you have seen is a front end — this is the three layers underneath it

A build guide for the part nobody records: the workspace agents run in, the files that tell them what the business is, and the access that lets them act without letting them send. Four layers, four paste-able prompts, and a build order that deliberately leaves the infrastructure until month two.

AgentsClaude CodeMCPSupabase+1read →
42ideatoolingSep 2026

Fifty sites, and the seven you will actually use

A bookmark list worth keeping — every domain checked, two had quietly moved

Fifty free web tools grouped by the job they do, rather than listed flat. All fifty were checked and every one is still up — but two have changed address since the list was written. The advice that makes it usable is in the list itself: do not try all fifty, pick the seven that match your week.

ToolsBookmarksFree toolsUtilitiesread →
43ideatoolingSep 2026

One key, twelve channels

Publishing to eleven social platforms from a single REST call — and the two costs that sit outside the flat monthly price

One POST replaces eleven platform APIs, each with its own OAuth flow, media rules and idea of what counts as a video. The scheduling call is the easy part — four lines. The parts worth knowing before you build on it are the two costs that sit outside the flat monthly price, and the fact that the field names in the curl examples are not the field names the SDKs put on the wire.

REST APISocial automationMCPScheduling+1read →
44ideamediaSep 2026

A shorts studio that runs in the terminal

One sentence in — script, visuals, voiceover, synced captions, sound design and a rendered vertical video out

A repo that turns Claude Code into a faceless-video studio. You describe the short in plain English and it writes the script, builds the visuals, generates the voiceover, syncs captions word-for-word, adds sound design and renders the vertical file. No camera, no face, no editing software — and on the default track, no per-clip generation cost either.

Claude CodeRemotionElevenLabsFaceless video+1read →
45ideatradingAug 2026

A Bloomberg terminal without the data feed

Six of the terminal’s seven layers on an agent swarm, with a chat bot as the cockpit

A Bloomberg seat bundles seven functions, and only one of them — the licensed real-time data — is actually protected by contracts. The rest are protected by workflow, which is the kind of moat agents dissolve. The architecture here is sound and the tools are real. The cost claim that headlines it is not: recomputed from its own inputs, the setup runs three to five times the price it advertises.

Kimi K3Grok BotAgent swarmDashboards+1read →
46ideatoolingAug 2026

Ten repos to check before you pay for a tool

Roughly 1.9M combined stars — directories, free tiers, local models and agent scaffolding

Ten repositories in the top tier of GitHub by stars that between them cover a lot of what a small team otherwise buys: an API directory, a free-tier directory, local model hosting, a scraper, a visual agent builder, an MCP index, and two agents that produce real files. Every star count here was checked live rather than copied.

Open sourceLocal LLMsMCPFree tiersread →
47ideamediaAug 2026

Codex CLI as a content pipeline, not a coding tool

One product photo in the terminal, listing images and ad videos out as files

What a coding agent is actually good at — read a goal, break it into steps, run tools, keep files organised — is the same shape as producing a launch campaign. This wires generation skills into Codex CLI so a single photo becomes listing images, UGC ads, commercials and thumbnails on disk.

Codex CLIAgent skillsGenerative mediaTerminal+1read →
48ideaagentsAug 2026

Two repos worth cloning before building an agent

Around 150 open-source agents and RAG apps — adapt one instead of starting from a blank project

Two GitHub repositories holding working agent projects across voice, customer support, research, RAG, coding and workflow automation. The point is not to read them — it is to find one close to what you need, open it in Claude Code, and describe your use case until it becomes yours.

AI agentsOpen sourceRAGMCP+2read →
49ideatradingAug 2026

An option-trading advisor built from Claude skills

25 skills and 32 MCP tools that turn plain English into quotes, Greeks, scans and portfolio reads

A set of Claude skills plus an MCP server that answer market questions with real data — option chains, Greeks, technicals, correlation matrices, PMCC scans, full PDF reports. Connect Interactive Brokers and it reads your live portfolio too. Everything but one skill is read-only, and it runs on the free Claude Desktop tier.

Claude SkillsMCPOptionsInteractive Brokers+2read →
50ideatradingAug 2026

100 mean-reversion strategies, as a working catalog

Futures, options and crypto — every entry, exit and stop, in a spreadsheet you can filter

A catalog of 100 systematic strategies weighted toward mean reversion, each with its precise entry, exit, stop, parameters and indicators. Eight of them were backtested on real data with costs on — those results are in their own tab, alongside the methodology for running the rest yourself.

Mean reversionFuturesOptionsCrypto+2read →
51ideatoolingAug 2026

The five skills to install before anything else

Memory across sessions, a brake on over-engineering, your notes as a knowledge base, and an image-to-video pipeline

Most people run Claude with zero skills installed, which is the difference between a chatbot and a system. These five are the ones worth setting up first — and the order matters, because each one makes the next more useful.

Claude CodeSkillsPluginsMemory+2read →
52ideaagentsAug 2026

The free agent web access setup

One open-source install that lets an agent read YouTube, GitHub, X and Reddit

Read and search access to the platforms an agent normally cannot touch — YouTube transcripts, GitHub, X posts, Reddit threads, any URL as clean Markdown — from one install rather than a stack of paid APIs. The X API alone runs about $215 a month at moderate use; this is MIT and free.

MCPClaude CodeResearchYouTube+4read →
53ideatoolingAug 2026

Five skills that turn Claude Code into an animation studio

Design DNA, GSAP, Three.js, Motion Design, Genjutsu — in the order that matters

A skill is not a plugin that adds a button — it is a manual Claude reads before writing. It stops inventing an API from memory and starts using the real one, pitfalls included. Five of them, each with its exact install command and the prompt to send immediately after. All MIT licensed.

Claude CodeSkillsGSAPThree.js+3read →
54ideamediaAug 2026

The Airbnb listing video system

Cinematic listing media for short-term rentals, made remotely from the host’s phone photos

Hosts pay thousands for professional listing photos and video. This generates both from the photos they already have on their phone — enhanced stills, a 20-second cinematic listing video, and a walkthrough tour — without ever visiting the property. $500 to $2,500 per property.

ClaudeMCPHiggsfieldGenerative video+2read →
55ideaagentsAug 2026

A multi-agent army on zero API keys

Three subscriptions, $240 a month flat — no meter, no per-token bill

Orchestration is the token bill, so the whole design is a delegation layer: expensive models decompose and judge, cheap models sprawl and summarise, and nothing raw ever crosses between them. The same work at API rates was tracking toward $15,000 over 10B tokens; on subscriptions it lands near $800.

Multi-agentClaude MaxOllama CloudCodex+2read →
56paper tradingtradingAug 2026

One setup, nothing else

A Claude Code agent that only trades reactions at decade-old price levels

Most trading bots lose because they take every setup they can see. This one is only allowed to do a single thing: wait for price to come back to a level that has mattered for years, and react there.

Claude CodeMCPRobinhoodSwing trading+1read →