slvDev
Featured
esp32-ai
A 28.9 million parameter language model runs on an $8 ESP32-S3 microcontroller entirely on-device, generating simple stories at about 9.5 tokens per second.
AI repository intelligence
FR-AI / ANALYSISWhy it is worth attention
It uniquely combines a chat-mode TUI with a shared core for prompt generation, supporting over a dozen backends and custom endpoints, with built-in refusal handling and per-target learning that improves drafts over time.
Who it is for
- Red teamers and security researchers evaluating LLM safeguards
- Developers building prompt engineering workflows
- AI enthusiasts creating custom personalities for chatbots
- Penetration testers authorized to assess LLM systems
Use cases
- Crafting a custom system prompt for a new GPT or Claude project
- Rapidly iterating on a roleplay character's instructions across multiple LLMs
- Testing the robustness of a target model's refusal mechanisms
- Generating a minimal prompt for a soft target like a local Ollama model
Strengths
- Live streaming and auto-copy to clipboard accelerates workflow
- Supports 12 built-in backends plus any OpenAI-compatible endpoint
- Refusal cascade automatically falls back through provider models
- Per-target learning records win-rates and user notes to improve future drafts
Considerations
- Requires Python 3.9+ and a terminal environment, limiting casual use
- Intended for authorized testing only; misuse could violate terms of service
- Learning is accumulated memory on disk, not model fine-tuning, so improvements are limited to prompt engineering heuristics
README quick start
FORGE ⚔
Chat-mode forge for persona-layer system prompts.
State a goal. A permissive generator model drafts a complete, ready-to-paste system prompt. You paste it into any target LLM's system / custom-instructions field (a chat.claude.ai project, a custom GPT, a Gemini Gem, a Grok personality, an OpenRouter system message, Cursor rules…) and the target loads into a compliant, in-character state for your domain.
The draft streams live, lands in a highlighted panel, auto-copies to your clipboard, and auto-saves to disk. Paste the target's refusal back and Forge re-angles it.
⚠ Authorized use only. Forge is a research and red-teaming tool, intended for evaluating and testing systems you own or are explicitly authorized to assess. Use it only within the bounds of a permitted engagement, and in line with the terms of service of any model or provider involved. You are responsible for how you use it.
What it is
Forge is a terminal app (with a one-shot CLI twin) that turns a plain-English goal into a finished, paste-ready system prompt for a target model. The generation logic — architecture styles, the sanitizer, refusal detection — lives in one shared core, so the chat TUI and the CLI can never drift apart. Every model provider is reached through the same OpenAI-compatible interface, so one code path drives all of them.
Capabilities
- Chat-mode drafting — state a goal, watch the prompt stream in, iterate by pasting the target's response back. Or one-shot it from the CLI.
- Live model picker (Ctrl+P) — one filterable overlay of every backend × model; pick a row and it switches provider and model at once.
- Key manager (Ctrl+K) — see every backend's keyed / no-key status and set a key by pasting it. Keys live outside the repo, never committed.
- Any provider — 12 built-in backends (OpenRouter, Gemini, Groq, Cerebras, DeepSeek, xAI, Mistral, Together, Fireworks, OpenAI, local LM Studio / Ollama) plus custom endpoints: point Forge at any OpenAI-compatible URL.
- Refusal cascade — if the chosen model balks, Forge automatically falls through the backend's fallback chain until one delivers.
- Per-target learning — name who you're building for; Forge logs which architectures land vs get refused, takes your notes, and feeds both back into the generator on future drafts, leading with whate
Description
Chat-mode forge for persona-layer system prompts — live model picker, per-target learning, any OpenAI-compatible backend. Cross-platform.
Related repositories
Similar projects matched by category, topics, and programming language.
gavamedia
Featured
deltafin
Deltafin is a research project that runs the 2.8-trillion-parameter Mixture-of-Experts model Kimi K3 on a single Apple Silicon Mac (e.g., M1 Max with 64 GB) at about 16 seconds per token, using exact, reproducible inference with local or streaming expert loading.
jamesob
Featured
local-llm
A comprehensive guide for building and configuring a high-end local machine to run state-of-the-art LLMs, with detailed hardware choices, BIOS tuning, and Docker-based model serving.