📦 mikiarlo3/ai-copywriter

AI Copywriter

Generate marketing copy that actually sounds human. This project, called AI copywriter, fixes the robotic tone that usually makes AI writing feel cold and generic. It works by combining real marketing skills with a strict set of rules designed to strip away every sign of artificial text. Before writing, it interviews you about your reader's specific feelings and your product's story. The result is punchy headlines and clear descriptions that feel like a person wrote them, not a machine. It is a simple tool that helps you connect with your audience using genuine language and real facts.

🆔 @hackernewsgithubprojects

📦 oniondas/awesome-hidden-ai-credits-free

10 Billion Free AI Tokens

One developer compiled a list that gives away ten billion free tokens every month, worth over one thousand six hundred dollars. This repository catalogs lesser-known platforms and APIs offering massive allowances, daily credits, and generous free tiers for text, images, and audio. It solves the problem of expensive compute costs by aggregating hidden resources into one place. You get access to top models through these free gates without paying for premium subscriptions. The maintainer updates the list weekly to ensure every link and limit remains accurate. It is a practical guide to building with advanced tools on a budget.

🆔 @hackernewsgithubprojects

📦 omdsh-dev/dsh-at-file

dsh-at-file: Point AI to Files Without Reading Them

The DeepSeek Harness plugin dsh-at-file lets you point an AI agent directly at specific workspace files using simple at-sign mentions, without manually pasting massive blocks of code into your chat. You type an at-sign in the composer, search for a file by name, and select it. The plugin then securely verifies that the file actually exists in your project before the agent starts working. Instead of sending the file's entire content to the AI, it just tags the file path as a reference.

🆔 @hackernewsgithubprojects

📦 henryqin1997/statem

StateM: The Runbook for Long AI Tasks

Keep AI agents on track during long tasks by storing their progress in a simple file-based runbook instead of relying on their memory. StateM works by turning complex workflows into a clear map of steps, where the agent can only move forward when it passes specific tests. This prevents the goal from getting lost or the agent from skipping important checks during a long session. It essentially acts as a durable note-taking system that survives even if the chat history is cleared. The result is a reliable way to handle complex work without needing a heavy workflow engine.

🆔 @hackernewsgithubprojects

📦 yunpeng1998/qwen-video-edit

Edit Video with an Image Model

Rewrite any video using a simple text command without training a brand new system. The project, Qwen Video Edit, takes a powerful image editing model and teaches it to handle video frames by treating them like a grid of pictures. It uses tiny, lightweight connections to bridge the gap between the image model and video data, allowing it to understand motion without expensive retraining. This means you can apply specific changes to long videos chunk by chunk, keeping everything consistent. It is a clever workaround that turns a single image tool into a full video editor.

🆔 @hackernewsgithubprojects

📦 csslc/pixrestore

PixRestore: The One-Step Image Fixer

PixRestore just dropped a model that fixes your blurry photos in a single pass, skipping the slow multi-step generation process that usually makes this take forever. Instead of relying on compressed latent spaces, it works directly on the raw pixels, which keeps the details crisp and avoids the weird artifacts you often see. It is built to handle a wide range of image problems in one go, so you do not need a different tool for noise or blur. The team also added a fast inference mode that lets it run quickly without sacrificing quality.

🆔 @hackernewsgithubprojects

📦 shapsider/pertmind

PertMind

Cellular perturbation databases can actually be turned into reinforcement learning environments where measured gene responses act as reward signals. A new project called pertmind is testing this very concept by utilizing experimental cellular outcomes to train biological language models. Instead of relying on manual training data, it uses public perturbation atlases to guide artificial intelligence toward structured biological reasoning. The system helps researchers predict how specific genes will react to custom cellular modifications or drug screens. By turning biological data into a learning playground, pertmind provides scientists with a smarter way to explore cellular behavior and generate fresh research hypotheses.

🆔 @hackernewsgithubprojects

📦 fresh-ma/harmprofile

The Hidden Map of AI Harm

Stop looking at AI safety as just a pass or fail grade. HarmProfile flips the script by mapping the actual harmful content models spit out, creating a distinct risk fingerprint for each system. Researchers collected over eighty thousand real examples from twenty-three major AI models, showing that while bigger models are better at talking, they often hide deeper, more dangerous knowledge underneath. It is like giving every AI a medical checkup based on its mistakes instead of just its good deeds. This dataset helps us see exactly where and how these systems go wrong.

🆔 @hackernewsgithubprojects

📦 startluxlabs/massive-activations-hla

Massive Activations HLA: The Hidden Rhythm of Efficient AI Models

Trace how massive activations spike precisely before full-attention layers in hybrid linear attention models, revealing a hidden rhythm that traditional Transformers completely miss. This repository gives you the official code and tools to reproduce these pre-attention spikes and inter-spike plateaus, showing exactly where and why these intense signals occur. By breaking down the lifecycle of these outliers, it proves that efficient hybrid models aren't just cheaper versions of standard architecture; they have their own distinct, predictable patterns of information flow.

🆔 @hackernewsgithubprojects

📦 jiyuan-tan/causalforge

CausalForge: AI Writes Math, Proofs Verify It

CausalForge lets an AI propose new theorems in causal inference, then forces them to be verified by a strict mathematical proof checker. This solves a major problem where machine learning models might guess wrong or hallucinate facts. Here, if the proof fails, the result is simply rejected. You get a library of mathematically certain causal rules, built by machines, without any human trust required. It is a fascinating intersection of logic and artificial intelligence. Watch how automated reasoning meets rigorous mathematics to create a self-improving system that actually works. Stay curious about the future of verified intelligence.

🆔 @hackernewsgithubprojects

📦 builderio/skills

Skills for Coding Agents

Skills is the toolkit that gives your AI coding agents real professional discipline. Instead of letting an AI guess or hallucinate outdated code patterns, this project equips them with specific, composable behaviors. The standout feature is the visual recap system, which turns messy code diffs into clear, interactive diagrams so you can actually understand the architecture changes before you merge. It also includes practical safeguards like checking official documentation before coding to prevent errors, and auditing one agent's work to ensure it actually finished the job. If you are using AI to write code, this collection turns unpredictable chat outputs into a reliable, review-friendly engineering workflow that you can trust.

🆔 @hackernewsgithubprojects

📦 galaxywalk/wave2body

Wave2Body: Seeing People Through Radar

Wave2Body lets you figure out exactly where a human’s limbs are in three dimensions using nothing but invisible millimeter wave radar signals. Instead of trying to guess coordinates from scratch, it learns to translate raw radar data into distinct body part tokens. This approach turns a complex math problem into a translation task, making the system significantly faster and more memory efficient. It’s a clever way to track movement without cameras, which is huge for privacy-friendly robotics. If you like smart hardware hacks, this repo is a must see.

🆔 @hackernewsgithubprojects

📦 trungdq88/mechanical-deployables-treejs

Mechanical Deployables

This repository lets you build 3D machines that rise from the earth, assembling part by part like heavy industrial machinery. It solves the common animation problem where objects just pop into existence, instead forcing every piece to telescope, hinge, or screw out of solid ground with realistic physics. The most fascinating feature is the destruction mode, where the entire structure collapses under gravity into a scattered wreck, proving the physics are real. You get eleven distinct models, from a simple watchtower to a massive colossus, all running smoothly in your browser.

🆔 @hackernewsgithubprojects

📦 tencent/ui-mate

Show It Once, Then Let It Work

Watch UI-Mate watch a human finish a complex desktop task once, then watch it adapt that workflow to do something slightly different without failing. Tencent built this AI to look at your live screen and click through apps just like you, but the real trick is learning from a single recorded example. Instead of needing a perfect text description for every job, you show it the workflow, and it figures out how to apply those steps to new situations. It’s not just replaying clicks; it’s understanding the logic behind them. This is a huge step toward AI that actually works with your specific tools and files.

🆔 @hackernewsgithubprojects

📦 spielewoy/autoprompt-skill

Autoprompt cuts AI coding errors by 45%

Autoprompt is a coding-agent skill that reduces failures by forty-five percent on agentic coding tasks. When you give an AI assistant a complex coding goal, Autoprompt takes over the execution loop. It breaks the job down, manages parallel work, and has independent agents review and verify the code before you sign off. The result is fewer mistakes and less back-and-forth. You still set the goal and constraints, but you don't have to prompt every step. It works with major AI coding tools like Claude Code and OpenCode. The trade-off is it takes about three times longer and uses twice the tokens, but for critical builds, that reliability is worth it.

🆔 @hackernewsgithubprojects

📦 denissergeevitch/desktop-fly

DesktopFly: A Virtual Fruit Fly With Real Neurons

DesktopFly is the desktop companion that puts a real brain on a tiny virtual insect. This project renders a 3D fruit fly on your Mac, but its decisions come from a live simulation of actual neural wiring found in real fruit flies. Instead of simple code that says if you click, it runs, the fly uses a circuit of real neurons to decide its moves. The coolest part is the escape reflex: when your mouse moves toward it, the fly’s brain actually processes the threat, and only takes off if its specific escape neurons fire.

📰 https://news.ycombinator.com/item?id=49353221

🆔 @hackernewsgithubprojects

📦 drozbay/maskvidexperiments

Fixing Video AI Jitter with MaskVidExperiments

MaskVidExperiments stops video AI from seeing shaky crops. When you zoom into a moving subject, the frame usually jumps around, tricking the model into thinking the camera is moving. This tool holds the crop steady, ignoring flickers so the AI only sees real motion. It then pastes the result back without any visible seams or jitter. If you work with video models, this is the fix for that annoying shake.

🆔 @hackernewsgithubprojects

📦 floe-labs/floe-guard

How Floe Guard Stops Runaway AI Costs

Floe Guard is the budget gate that stops your AI voice agents before they burn through your money. Most cost trackers tell you what you spent after the fact, but this tool sits right in the middle of the conversation to hard stop the next turn if it would cross your ceiling. It works locally without needing a cloud account or sending data to anyone. The coolest part is that it prices the entire voice call, including speech recognition, text generation, and telephony, so you know the real cost of a single interaction.

🆔 @hackernewsgithubprojects

📦 vercel-labs/fx

The Coding Agent That Feels Like a Unix Shell

fx is the terminal coding agent that finally feels like a standard Unix shell instead of a heavy application. Built entirely in Zig, it compiles down to a tiny eight megabyte binary, which makes it insanely fast and easy to drop into other systems without bloat. It runs locally or in the cloud, works with any AI model you prefer, and stays out of your way. The real kicker is how clean the interface is; no clunky widgets, just a simple prompt where you can actually read the code changes and commands running in real time. It is open source, experimental, and built for people who love their command line.

📰 https://news.ycombinator.com/item?id=49353799

🆔 @hackernewsgithubprojects

📦 ygtq3/disk-butler

Disk Butler

Disk Butler is a tiny, two-point-four-megabyte Windows tool that scans your drive in seconds, not minutes. It reads the disk's master file table directly, so it sees every file instantly without slowing your computer. The interface uses plain language to explain what each folder is and whether it is safe to delete, removing the guesswork and fear from cleaning up your system. It runs entirely offline, meaning your data never leaves your machine. Instead of scary pop-ups or hidden services, it gives you clear, honest control over your digital space.

🆔 @hackernewsgithubprojects

📦 ariescar/anycreature

AnyCreature: AI-Powered 3D Beast Builder

AnyCreature is the text-to-3D pipeline that turns a simple written order into a fully animated, game-ready model. This is the creature generator that actually works, solving the nightmare of hand-sculpting every asset from scratch. Instead of downloading generic packs or wrestling with complex software, you just type a prompt like "menacing mountain giant." The system handles the heavy lifting, asking only a couple of clarifying questions before it builds the shape, colors, and animations itself. The coolest part is its strict quality control, where a separate AI agent acts as a blind judge to ensure the final model actually looks good, not just technically correct.

📰 https://news.ycombinator.com/item?id=49343561

🆔 @hackernewsgithubprojects