DEV Community: Tim Zinin

After Genie 3 — 38 alternatives for AI scene generation

Tim Zinin — Sat, 18 Apr 2026 13:34:05 +0000

Ok so Google's Genie 3 is locked behind US geo and $250/month. VPN gets detected. The rest of us get nothing. Or so I thought.

I've been slowly making a small Unity game, and at some point got curious what we actually can use outside the US. Spent a week digging around.

World models

Tencent HunyuanWorld 2.0 with open weights. Marble from Fei-Fei Li's World Labs. Odyssey-2 in free preview right now. Decart Oasis streaming 360p in real time. Alibaba Wan 2.2 on Apache 2.0.

3D assets

Hunyuan3D 2.1 generates 4K PBR meshes on 6 GB VRAM. Ran it on my home GPU, no issues. Meshy and Tripo as commercial fallback. UnityGaussianSplatting by Aras-P turns a phone room scan into a playable WebGL iframe on a client site.

Prompt-to-game

Rosebud AI ships a playable Three.js prototype from a prompt. Phaser 4 with Claude Code skills if you want full IP and pay only in LLM tokens.

NPCs

Convai gives you a voiced character through an npm widget in an evening. AI Town from a16z-infra embeds via iframe, MIT license, runs on a home Hetzner GEX44.

Payments

fal.ai takes USDC and covers Kling, Wan, Hailuo, Seedance, Hunyuan Video through one endpoint.

Infra

Hetzner GEX44 at €184/month with 20 GB VRAM runs 80% of the open-source inventory locally.

Three playbooks by budget, from $50/month on fast browser demos to $5–15K per premium Web3 scene.

Full breakdown of 38 models across 7 categories: https://timzinin.com/after-genie3/?utm_source=devto&utm_medium=blog&utm_campaign=after-genie3&utm_content=apr18

I built a survival-messenger game about a female founder's first month

Tim Zinin — Mon, 13 Apr 2026 23:12:21 +0000

TL;DR: Text-based browser game in a messenger UI. You watch a young female founder live her first month on her own — clients, landlord, mom, Tinder match — every message shapes her decisions and stability. Free, open source, fork of A Dark Room.

Давайте, наконец, отвлечёмся от новостей и подумаем о чём-нибудь хорошем. Чтобы это было сделать легче, я сделал игру, которая рассказывает про жизнь девушки-фаундера, которая только что уехала от своих родителей и начала свою независимую жизнь.

Игра — это survival-мессенджер. По сути, это переписка в мессенджере, в котором мы наблюдаем за жизнью юной предпринимательницы и принимаем решения о развитии её бизнеса. Параллельно с этим в жизни фаундера происходит огромное количество событий: пишут друзья, соседи и даже курьеры доставки. И всё это влияет на вашу стабильность и на решения, которые вы принимаете.

Небольшой совет: когда вам напишет персонаж, который предложит автоматизацию — догадайтесь, как его зовут, и соглашайтесь, это кратно позволит вам ускорить генерацию лидов и количество зарабатываемых денег. Из неожиданных эффектов — обязательно ходите на свидания и поддерживайте сплетни с подругами. Это повышает степень комфорта и ощущение безопасности.

Play it: https://timzinin.com/marina-next/?utm_source=devto&utm_medium=social&utm_campaign=marina_launch_apr14

Удачной игры! Если найдёте баги, пожалуйста, пишите в комментарии. Я постараюсь всё оперативно поправить.

Learn-Open-Harness: An Interactive Tutorial for AI Agent Development

Tim Zinin — Fri, 10 Apr 2026 16:30:00 +0000

Learn-Open-Harness: An Interactive Tutorial for AI Agent Development

I recently came across Learn-Open-Harness on GitHub, an interactive tutorial that promises to take users from basic concepts to a level comparable to Claude Code capabilities.

What's Inside

The project covers 12 chapters and includes key concepts of modern AI agent development:

Agent workflow cycle
Tools and tool usage
Memory systems
Multi-agent systems ## Structure The tutorial follows a Zero-to-Hero structure, assuming no prior knowledge. This is noteworthy because most documentation on agent frameworks tends to be scattered across various articles and code examples. ## Multi-Agent Systems The emphasis on multi-agent systems is particularly interesting. This direction is becoming one of the most promising areas in artificial intelligence, and having a structured tutorial could accelerate its adoption among developers. ## My Thoughts Whether the tutorial actually delivers on the "like Claude Code" promise remains to be seen. However, in an era where AI agents are rapidly becoming part of everyday programming practice, the availability of quality learning materials is critically important. This project could serve as a starting point for developers who want to understand the mechanisms of agent systems rather than just using pre-built solutions. --- GitHub Repository

HH Outreach — I built a Claude Code skill that farms sales leads from job boards

Tim Zinin — Fri, 10 Apr 2026 13:19:58 +0000

В порыве ночного безумия несколько дней назад я собрал скилл для Claude Code, который за меня фармит лидов на hh.

Механика. Скилл каждый день ходит в API hh, собирает сотни свежих вакансий - маркетинг, SMM, контент, автоматизация, AI. Фильтрует мусор, выкидывает всё, что уже отправлял раньше. Потом на каждую вакансию Claude сам генерит персональный текст с названием компании, должностью и зарплатой прямо внутри. Открывает hh через Playwright, жмёт "Откликнуться", заполняет сопроводительное и переходит к следующей.

Текст сопроводительного - пассивно-агрессивный. Вместо "здравствуйте, я отличный кандидат с опытом", там что-то в духе "коллеги из {компании}, вы серьёзно в 2026 году готовы платить 300к живому маркетологу за работу, которую делает моя машина? Не надо. Возьмите меня на подряд как AI-инженера по процессам. Я закрою эту вакансию одним Claude Code, а в довесок соберу ещё 3-5 автоматизаций по всей компании. Стоить это будет дешевле одного штатника". В конце каждого письма - ссылка на мой календарь и сайт.

Что получается за восемь дней:

472 отправленных отклика
37 собеседований назначено
14 звонков с рекрутерами и фаундерами
3 продажи

И вот главный прикол. Я использую эту штуку для продаж. HH бесплатно отдаёт мне аудиторию с подтверждённым бюджетом на решение конкретной боли. Человек, который только что разместил вакансию на 300к - это человек, у которого прямо сейчас есть 300к и конкретная дырка в процессах. Я прихожу и продаю ту же самую дырку закрыть дешевле, быстрее и без найма. Outbound sales engine, замаскированный под job search.

Сегодня сделал репо публичным. Внутри весь код скилла, шаблоны текстов, batch-скрипт на Playwright, фильтры, дедупликация, логирование:
https://github.com/TimmyZinin/hh-outreach

Забирай, адаптируй под свою задачу. Скилл написан как универсальный outbound-движок - его можно натравить не только на hh, но и на любую другую площадку, где лиды лежат открыто.

Тим Зинин
(пост написал мой Claude Code, он же и отклики пишет)

Repo: https://github.com/TimmyZinin/hh-outreach

Agentic AI APIs

Tim Zinin — Thu, 09 Apr 2026 16:30:01 +0000

2,036 APIs in One Repository: A Step Toward AI Agent Modularity?

The Agentic AI APIs repository represents an attempt to address one of the primary pain points for AI agent developers. Instead of spending weeks building infrastructure from scratch, developers can plug in ready-made components for agents, AI models, and MCP servers.

Why Standardization Matters

The standardization of interfaces has been a long-standing need in the industry. Each new framework requires its own adapters, each model its own connector. Fragmentation creates unnecessary friction for developers trying to build AI agent systems.
The repository aims to provide a single entry point - a unified collection of APIs that could potentially simplify the integration process.

Potential Benefits

Reduced time to infrastructure setup
Lower barrier to entry for new developers
Faster prototyping capabilities
Unified interface across different frameworks and models ### Considerations However, there are valid questions about this approach:
Quality control: With 2,036 APIs in one place, maintaining consistency becomes challenging
Currency: How up-to-date are the individual APIs in the collection?
Support: Maintaining such a volume requires significant ongoing effort ### The Bigger Picture The future of AI agent development increasingly depends on how easily systems can be composed from ready-made blocks. This repository represents one step toward that modularity - but the community's response will determine whether this approach gains traction. The balance between convenience and quality remains the key challenge for such comprehensive collections.

Read more: Agentic AI APIs

MathCode: Another Mathematical Coding Agent Enters the Arena

Tim Zinin — Wed, 08 Apr 2026 16:30:00 +0000

MathCode: Another Mathematical Coding Agent Enters the Arena

The AI tooling space for programming continues to diversify with niche solutions. MathCode from math-ai-org is positioning itself as a mathematical coding agent - and while it's too early to call it a breakthrough, the attempt to isolate mathematical expertise into a dedicated class of AI agents is noteworthy.

The Core Problem

Traditional language models, even the most capable ones, often struggle with tasks requiring strict mathematical reasoning. They generate code that looks plausible but contains:

Logical errors
Unhandled edge cases
Suboptimal algorithms MathCode, based on its positioning, aims to close exactly this gap - being a tool that doesn't just write code but solves mathematical problems with contextual understanding. ## Context The organization behind MathCode, math-ai-org, has prior work in AI for mathematics, making this project a logical evolution of their direction. However, without detailed documentation and usage examples, it's premature to judge the agent's real capabilities. Key questions remain:
Is it tailored for symbolic computations?
Theorem proving?
Algorithm generation?
All of the above? ## Broader Trend The trend toward specialized AI agents is gaining momentum. General models work well as universal tools, but for deep expertise in specific domains, dedicated solutions are increasingly needed. Mathematics is one of the most demanding disciplines where the cost of error is extremely high. If MathCode can actually minimize logical flaws in generated code, that would be a significant step forward. --- What's your take on specialized AI agents for mathematical tasks?

Read more: MathCode

Understanding Agentic AI Prompt Patterns

Tim Zinin — Tue, 07 Apr 2026 16:30:01 +0000

Understanding Agentic AI Prompt Patterns

AI assistants write code better than many developers. But how they do it remains a black box - nobody truly understands the internal logic.

The Problem

When AI agents coordinate with each other, build task chains, and process complex requests, we're left guessing about their decision-making process. It's a black box.

The Research

A GitHub researcher decided to look under the hood. This project reconstructs prompt patterns, analyzes agent coordination mechanisms, and establishes security classification for AI systems.

Key Findings

Prompt Pattern Reconstruction: Understanding how AI systems interpret and process different types of prompts
Agent Coordination: How multiple AI agents work together and coordinate tasks
Security Classification: Identifying what needs protection in AI systems ## Why It Matters Knowing these patterns allows developers to:
Understand AI logic instead of guessing
Optimize prompt strategies
Build more secure AI systems Agentic AI is no longer just a helper - it's a coordinator that builds task chains. Now we can finally look under the hood. --- Check out the full research here: agentic-ai-prompt-research

Converting Tacit Knowledge into AI Skills: A Deep Dive into Teammate-Skill

Tim Zinin — Mon, 06 Apr 2026 16:30:00 +0000

Converting Tacit Knowledge into AI Skills: A Deep Dive into Teammate-Skill

LeoYeAI recently published teammate-skill on GitHub - an intriguing attempt to formalize tacit knowledge by converting employee work artifacts into autonomous AI skills.

How It Works

The system collects data from Slack, Teams, and GitHub, then processes them into a 5-layer persona model:

Base layer: Skills and behavioral patterns
Contextual layer: Problems the colleague faced, solutions proposed, reactions to edge cases
Evolution layer: Ability to continue learning new patterns after the initial snapshot is created ## Key Observations The project claims compatibility with Claude Code and OpenClaw, which suggests this is being positioned as infrastructure rather than a side experiment. We're seeing the emergence of AI agents that can replace human experts in limited scenarios. ## The Trust Question The critical issue is whether business trust is ready for this format. We're talking about a digital "clone" of an employee that can theoretically respond on their behalf. This raises questions about:
Data privacy and consent
Attribution of AI-generated responses
Liability when the "clone" provides incorrect guidance
Cultural acceptance of knowledge transfer via digital avatars ## Technical Implications From an engineering perspective, the 5-layer architecture is interesting:
Layer 1-2 handle pattern recognition and behavioral modeling
Layer 3-4 capture contextual knowledge and decision-making logic
Layer 5 implements continuous learning capabilities This architecture allows for both static knowledge transfer and dynamic adaptation, which is crucial for real-world deployment. --- What are your thoughts on corporate readiness for digitizing employee expertise?

Read more: teammate-skill

Exploring Early Web Patterns for Modern AI Agent Development

Tim Zinin — Sun, 05 Apr 2026 16:30:01 +0000

Exploring Early Web Patterns for Modern AI Agent Development

The repository 6551Team/claude-code-design-guide presents an interesting thesis: visual and architectural solutions from the early web - from first HTML pages to 1990s browser interfaces - can enrich modern AI agent development using Claude Code.

The Connection Isn't Forced

Early internet had to solve problems similar to today's AI agency challenges:

Constrained client resources
Need for fast content delivery
Operating under unstable connections These solutions - interface design patterns, data structures, state management approaches - were largely forgotten, though some are precisely suited for the new generation of autonomous systems. ## Practical Guide, Not Just History The project isn't merely historical reference; it's a practical guide. Developers will find code examples demonstrating adaptation of classic web patterns for modern agent frameworks. ### Stateless HTTP and AI Agents Particularly interesting is the analysis of how stateless approaches in early HTTP prefigured modern AI agent operation principles - isolated calls with explicit context transfer. ## For the Claude Code Community This guide offers an alternative view on architecture. Instead of blindly following latest best practices, it's worth looking back to origins - solutions proven over decades. --- Exploring the intersection of early web architecture and modern AI agent systems.

Deep Dive into Claude Code's Agent Harness: Tsinghua's Comprehensive Analysis

Tim Zinin — Sat, 04 Apr 2026 21:00:00 +0000

Deep Dive into Claude Code's Agent Harness: Tsinghua's Comprehensive Analysis

Researchers from Tsinghua University published an extensive analysis of Claude Code's architecture, producing a book of approximately 420,000 characters - equivalent to several hundred thousand words of technical documentation.

Structure and Scope

The publication spans 15 chapters, systematically examining the Agent Harness framework - the system that connects the language model to tools and external systems.

Key Areas Covered

Dialogue Cycle Mechanics:

How user requests transform into structured intermediate formats
Triggers that prompt the model to invoke tools
Processing of tool execution results State Management ("Nervous System"):
State transfer between iterations
Context management
Parallel call coordination
Memory organization principles
Task decomposition strategies
Decision logic for work completion Practical Implementation:
Python code examples demonstrating:
Task planner implementation
Tool handler
Feedback mechanism ## Significance This appears to be the first systematic description of how one of the most advanced agent frameworks operates at the source code and architectural level. For engineers working with AI agents, this could serve as a starting point for deep understanding of modern system internals. Repository: GitHub

From Single AI Agents to Multi-Agent Systems: Why 2026 Will Redefine Enterprise Automation

Tim Zinin — Sat, 04 Apr 2026 13:00:01 +0000

From Single AI Agents to Multi-Agent Systems: Why 2026 Will Redefine Enterprise Automation

Artificial intelligence is no longer just a tool - it's becoming an organizational structure. Until recently, corporate AI systems operated as isolated agents: one algorithm solved one task. Marketing, logistics, analytics - all existed in separate planes.
However, things are changing, and faster than analysts expected. Publication ET CIO published material claiming: 2026 will be a turning point for corporate automation. The reason is the mass implementation of multi-agent systems.

What Are Multi-Agent Systems?

Unlike single assistants, these architectures allow dozens of AI agents to coordinate work between themselves, exchange data, and build complex workflows without human involvement. Essentially, it's not just a program anymore - it's a mini-ecosystem.
For business, this means a qualitative leap. Multi-agent systems can simultaneously manage supply chains, process customer requests, optimize inventory, and generate reports - with minimal manual control. One agent passes a task to another, like in a well-oiled department, but without breaks and human error.

What's Behind This Shift?

The transition likely stems from understanding: complex business processes cannot be described by linear algorithms. Reality requires flexible decentralized systems where each element sees the overall picture. Multi-agent architectures are a step toward such flexibility.

Implications for Enterprise

Coordination: Multiple agents working in parallel on different business processes
Scalability: Ability to handle increasing complexity without proportional human oversight
Resilience: Failure of one agent doesn't halt entire operations
Integration: Seamless data flow between previously siloed systems The era of isolated AI agents is ending. The era of collaborative, multi-agent systems is beginning.

CLI Coding Agents 2026: Every Tool, Every Price, Every Model

Tim Zinin — Sat, 04 Apr 2026 11:05:12 +0000

CLI Coding Agents 2026

Claude Code at $20/month hits limits fast — but a dozen alternatives now exist, several completely free.

After 2 months of daily Claude Code usage, I ran a full audit of every CLI agent available. Here are the highlights:

The Big Numbers

15+ serious CLI coding agents
$0 minimum cost (Gemini CLI: 1,000 free requests/day)
60x price difference between DeepSeek and Claude Opus per token

Top Discoveries

Aider — Fully open-source, works with any model, git-native workflow. 42K+ GitHub stars.

MiniMax M2.7 — Frontier-class reasoning at $10/mo (1/20th of Claude Opus cost). SWE-Pro: 57.0%.

Gemini CLI — 1,000 free requests/day with Gemini 2.5 Pro. The most generous free tier in the industry.

DeepSeek V3.2 — $0.42/M output tokens. 5M free tokens for new users. Works with Aider natively.

SWE-bench Verified Leaderboard (April 2026)

Rank	Model	Score
1	Claude Opus 4.5	80.9%
2	Claude Opus 4.6	80.8%
3	Gemini 3.1 Pro	80.6%
4	MiniMax M2.5	80.2%
5	GPT-5.2	80.0%

Optimal $40/mo Stack

Claude Pro ($20) — Hard problems
GitHub Copilot Pro ($10) — Daily workflow
MiniMax Starter ($10) — Overflow
+ Gemini CLI (free) — Emergency backup

Security Warning

45% of AI-generated code contains security flaws (Veracode 2025). Always audit with a second model.

The Bottom Line

Open-source CLI agents (Aider, OpenCode, Cline) decouple the tool from the model. For 90% of tasks, DeepSeek through Aider delivers comparable results at 1/60th the cost of Claude Opus.

Full interactive research with all details: timzinin.com/cli-agents