bggoranoff/replaceme

Última actividad de origen registrada: 26 de octubre de 2026 a las 12:27

Catálogo de SkillsMP actualizado: 2 de noviembre de 2026 a las 18:02

skills recopiladas: 37

Estrellas en GitHub: 0
Forks en GitHub: 0

Instalar con un asistente de IA

Copia este prompt en el asistente de IA que estés usando.

Follow https://skillsmp.com/skill-install/prompt.md to install Agent Skills from https://github.com/bggoranoff/replaceme.

README del repositorio

Archivo de origen del README: README.md

Leer el README original

Mostrar README completo

# ReplaceMe **Type a job. Get a game that teaches it. Watch an AI agent learn it by playing, and keep the skills it writes down.** [![License: MIT](https://img.shields.io/badge/License-MIT-yellow.svg)](LICENSE) ![Python 3.11+](https://img.shields.io/badge/python-3.11%2B-blue) ![Node 22+](https://img.shields.io/badge/node-22%2B-green) ![Built at {Tech: Europe} Agentic AI Hack](https://img.shields.io/badge/built%20at-%7BTech%3A%20Europe%7D%20Agentic%20AI%20Hack-black) ReplaceMe turns any hands-on job (barista, florist, doctor, rocket scientist) into a cozy, Stardew-style village game. A researcher agent studies the job on the web, a builder agent turns that research into four playable quest levels with fixed scoring rules, and an actor agent plays those levels like a trainee would. It changes one thing at a time, reads the feedback, and exports one reusable **skill** per level: a `SKILL.md` with the learned parameters, the evidence behind them, and the steps to run them with real tools. > Built in one day at the [**{Tech: Europe} Agentic AI Hack**](https://luma.com/zrh-hack?tk=K2XLTF) in Zürich (October 2026), hosted with Google DeepMind, with Gradium and Entire as technology partners. ## Contents - [How it works](#how-it-works) - [Features](#features) - [Quick start](#quick-start) - [Model providers](#model-providers) - [Project structure](#project-structure) - [Testing](#testing) - [Limitations](#limitations) - [Team](#team) - [License](#license) ## How it works ```mermaid flowchart LR U([Job prompt typed or spoken]) --> R[Researcher web-grounded brief] R -->|research.json| B[Builder village + 4 levels] B -->|world.json on branch game/slug| G[Game React + canvas] G <-->|same actions as humans| A[Actor learns by playing] A --> S[(skills/<level>/ SKILL.md + skill.json)] S --> C[Skill chat ask and refine] ``` 1. **Research.** The researcher picks five focused tasks for the job, then runs one web-grounded report per task. Each report cites its numbers: variables, units, optimal ranges, ratios, and what goes wrong when you are too high or too low. A validator checks every task: bounds, cited optima, reachable slider steps, and no answers leaked into the hints. It repairs what it can and drops anything unplayable. 2. **Build.** The builder generates one procedurally laid out village with quest-giver NPCs. It maps four tasks onto the platform's five mechanics (`mix`, `timing`, `sort`, `deliver`, `converse`) and writes declarative, deterministic scoring, mirrored in Python and TypeScript. A rules check runs against `main` before the world is committed to its own `game/<slug>` git branch. 3. **Play.** Humans walk the map with the arrow keys, take quests and play the minigames. Every level ends with a score card that gives qualitative feedback and never shows the targets. 4. **Train.** **Train agent** starts the actor. It drives headless Chromium through the same allowlisted controls a human uses, four levels in parallel. Each attempt is planned from the evidence so far: bracketing and coordinate search on sub-scores, BFS for delivery routes, and an LLM that reflects on the feedback. It never reads the scoring targets. 5. **Skills.** Each level exports a skill with learned values, control effects, failure modes, attempt evidence and tool mappings. You can chat with a skill, and telling it about your own context (equipment, climate, customers) updates the skill. Example output: [`games/barista/training/skills/barista-brew-coffee/SKILL.md`](games/barista/training/skills/barista-brew-coffee/SKILL.md). That skill learned an 18 g / 300 ml / 93 °C pour-over with a 3:00 to 3:30 brew window, with every value backed by attempt scores. ## Features - **Job-agnostic.** No code knows about any particular job. Everything comes from the prompt and the sources found. - **Four model providers.** Google Gemini, Anthropic Claude, OpenAI, or OpenAI models on AWS Bedrock, chosen with one environment variable. - **Deterministic and auditable.** Scoring is fixed and declarative, so model output can never overwrite game state. Reasoning traces are stored in [Entire](https://entire.io) Brain, and every game lives on its own git branch. - **Works offline.** Without a key, prebuilt games load, the actor falls back to pure deterministic search, and skill chat answers from the training evidence. - **Voice.** Optional speech input and spoken NPC lines via [Gradium](https://gradium.ai). - **Spectator mode.** Watch live training runs and switch between them with Left/Right. ## Quick start Requirements: Python 3.11+, Node 22+, git. ```bash git clone https://github.com/bggoranoff/replaceme.git cd replaceme python3 -m venv .venv && source .venv/bin/activate python3 -m pip install -r requirements.txt python3 -m playwright install chromium (cd frontend && npm install) cp .env.example .env # add one API key, see below cd frontend && npm run dev # open http://localhost:5173 ``` Type `barista` to load the prebuilt game instantly, or any other job to research and build a new one (about 2 to 4 minutes). For a step-by-step guide with checks and troubleshooting, see [`.claude/skills/run-locally/SKILL.md`](.claude/skills/run-locally/SKILL.md). [Claude Code](https://claude.com/claude-code) also picks that file up as a skill: ask it to "run ReplaceMe locally". ## Model providers Set one key in `.env`. `LLM_PROVIDER` picks a provider explicitly. When it is empty, the first provider with a key is used. `LLM_MODEL` overrides the default model. | Provider | Key | Default model | | --- | --- | --- | | Google Gemini | `GEMINI_API_KEY` | `gemini-3.8-flash` | | Anthropic | `ANTHROPIC_API_KEY` | `claude-opus-5-5` | | OpenAI | `OPENAI_API_KEY` | `gpt-6-luna` | | AWS Bedrock | `AWS_BEARER_TOKEN_BEDROCK` + `AWS_REGION` | `openai.gpt-6-luna` (via `bedrock-mantle`) | Researching new jobs needs a provider with web search, because uncited tasks are dropped. Playing, training and skill chat work with any of them. | Variable | Purpose | | --- | --- | | `GRADIUM_API_KEY` | Speech input and TTS (optional) | | `PREBUILT_GAMES=0` | Ignore prebuilt games and cached trainings, so everything runs live | | `ACTOR_ATTEMPTS` | Attempts per level (default 10) | | `ACTOR_USE_GEMINI=0` | Actor uses deterministic search only, with no model calls | ## Project structure ``` researcher/ job -> web-grounded research brief (research.json) builder/ brief -> world.json, deterministic scoring, rules check, model client actor/ plays levels in headless Chromium, learns, exports skills shared/ contracts: actions.json, schemas, Entire Brain helpers frontend/ React + Vite app; server/ holds the /api middleware that runs the agents games/ generated games (barista is prebuilt, with a recorded training) skills/ exported skills (SKILL.md + skill.json per level) fixtures/ example handoffs used by the tests scripts/ test suites and maintenance scripts ``` The contracts between the components are in [`shared/game_contract.md`](shared/game_contract.md). Each agent has its own README: [researcher](researcher/README.md), [builder](builder/README.md), [actor](actor/README.md). ## Testing ```bash for t in check_contracts test_world_builder test_brain test_builder_brain test_builder_contract test_graph_check test_actor; do python3 scripts/$t.py || echo "FAILED: $t" done (cd frontend && npm run build && node --experimental-strip-types --test server/skills/*.test.ts server/api/*.test.ts) ``` All suites run offline, except the live Gradium round trip, which needs a valid `GRADIUM_API_KEY`. ## Limitations - Simulation success is not real-world competence. Physical steps are simulated actions, and exported skills need matching tool or device adapters plus real validation before anyone relies on them. - This is a hackathon prototype: a single dev server, local JSON storage, and no auth. - `deploy.sh` and [`PLATFORMS.md`](PLATFORMS.md) describe the event's Google Cloud setup and need your own project to reuse. ## Team Built by [Boris Goranov](https://github.com/bggoranoff), Kaloyan Fachikov and Nikola Staykov at the {Tech: Europe} Agentic AI Hack, Zürich. ## License [MIT](LICENSE)

Cargando el README del repositorio…

Skills en este repositorio

clasificación pendiente

skill

ocupación

descripción

actualizado

run-locally

sin clasificar

Set up and run ReplaceMe on this machine with a Google Gemini, Anthropic, OpenAI or AWS Bedrock API key, then check that it works. Use when someone wants to install, configure, start, reproduce or demo the project locally, or switch the model provider.