OpenClaw: Nine Best Practices from the World’s Largest Vibe Coding Project

“I used to lead a team. I had a lot of software engineers under me. Even then, I had to accept that they would not write code exactly the way I wanted it written.” – Peter Steinberger¹

OpenClaw is the most-starred software repository on GitHub (310K+ stars) and the definitive case study in large-scale AI-assisted “vibe coding.”² Built by Austrian developer Peter Steinberger using three to eight parallel AI agent instances,³ the project demonstrates that a single developer can orchestrate a 300K+ LOC TypeScript monorepo, more than 20 messaging integrations, and native apps for three platforms — while barely reading most of the code.¹

The project’s name went through a turbulent journey. It began in November 2025 as Clawdbot, a prototype built in an hour on top of Anthropic’s Claude. After going public on GitHub in January 2026, it exploded to 9,000 stars in a single day, but a trademark warning from Anthropic forced a rename to Moltbot. Then, after security incidents involving impersonation accounts and malicious npm packages, it was renamed again just three days later to OpenClaw.⁴ In February 2026, Steinberger joined OpenAI to lead “the next generation of personal agents,” and OpenClaw was transferred to an independent foundation backed by OpenAI.⁵

Ironically, Steinberger himself calls the phrase “vibe coding” a slur and prefers “agentic engineering,”⁶ yet OpenClaw became the flagship project of that very movement.⁷ This post analyzes OpenClaw’s repository structure, workflows, and community operations, and distills the actionable best practices found there.

TL;DR

• OpenClaw is the definitive vibe coding case in which a single developer coordinated multiple AI agents to run a large TypeScript monorepo and native apps.
• The core patterns are a living AGENTS.md, clear PR norms, automated quality gates, and rules for coordinating parallel agents.
• Mastery in vibe coding shifts away from the ability to write a lot of code yourself, toward the ability to design and control the system that writes the code.

1. What OpenClaw Actually Does

OpenClaw is a self-hosted personal AI assistant that connects messaging platforms to large language models. Unlike the web interfaces of ChatGPT or Claude, OpenClaw runs on the user’s local machine and plugs into the channels the user already inhabits. It supports more than 15 channels, including WhatsApp, Telegram, Discord, Slack, Signal, iMessage, Microsoft Teams, Matrix, and LINE. The README puts it simply: “A personal AI assistant that runs right on your device.”⁸

The heart of the architecture is a local Gateway daemon. It is a WebSocket-based control plane running on port 18789 that routes incoming messages to independent AI agents. Each agent has its own workspace, memory, and personality defined by Markdown files (SOUL.md, MEMORY.md, USER.md).⁹ The agents do more than talk. They run shell commands, control browsers via CDP, manage schedules, execute cron jobs, and even reach out proactively through a heartbeat system.¹⁰ A skills marketplace called ClawHub hosts more than 1,700 community-built extensions.¹¹

The tech stack is a TypeScript-based pnpm monorepo (Node.js 22+), plus native companion apps written in Swift (macOS/iOS) and Kotlin (Android). Testing runs on Vitest with a 70%+ V8 coverage threshold.¹²

2. Traces of AI-Assisted Development Visible in the Repository

The clearest evidence of vibe coding lives in the commit history. Because early development used Anthropic’s Claude Code, “claude” appears frequently as a co-author on commits. After Steinberger joined OpenAI in February 2026, however, branches autonomously generated by the OpenAI Codex agent — names like codex/issue-issue-41258-20260312044119 — increased markedly. A single repository now carries traces of two AI coding agents, Claude and Codex, side by side. The README states it outright: “AI/vibe-coded PRs welcome!”⁸

Steinberger’s personal workflow is especially striking. He runs three to eight Codex CLI instances simultaneously in a 3x3 terminal grid, and most of them work in the same folder rather than in separate worktrees.³ Each agent makes atomic commits guided by the rules in AGENTS.md. He landed more than 6,600 commits in January 2026 alone — outwardly the pace of a 20-person team, but in reality one person combined with AI agents.¹ His prompts have grown shorter over time; these days one or two sentences plus a single screenshot is usually enough, with screenshots accounting for about 50% of the input.¹³

CONTRIBUTING.md codifies how the community should handle AI-assisted PRs.¹⁴

• Disclose AI usage in the PR title or description
• State the level of testing (untested / lightly tested / thoroughly tested)
• Include the prompt or session log where possible
• Confirm that you understand what the generated code does

The document closes this way: “We treat AI PRs as first-class citizens here. We just want transparency so reviewers know where to focus.”¹⁴

3. The AGENTS.md and CLAUDE.md Configuration Pattern

The most reproducible innovation in the OpenClaw repository is the AGENTS.md file.¹² It is a comprehensive instruction document for every AI coding agent working in the codebase. CLAUDE.md is a symlink pointing to the same file, guaranteeing that Claude Code and Codex-family agents read identical instructions. The rule is explicit: “When you add a new AGENTS.md anywhere in the repo, you must also add a CLAUDE.md symlink pointing to it.”¹²

The file operates as organizational memory for AI agents. Steinberger describes it as “a collection of organizational scar tissue,” because Codex itself has incrementally added content every time something went wrong.¹³ Its key sections:

• Build and test commands: Spelled out precisely — pnpm build, pnpm check, pnpm test, and pnpm tsgo for fast type checking
• Git conventions: Conventional Commits enforced with subsystem scopes such as fix(telegram): and feat(skills):, plus a required git pull --rebase before pushing
• Multi-agent safety rules: Never create, apply, or drop a git stash; never switch branches unless asked; if you find changes you do not recognize, assume another agent is working and carry on
• Changelog rules: Add only user-meaningful entries at the end of the section, and credit external contributors with the Thanks @author pattern
• Security boundaries: Never commit real phone numbers, production configuration values, or video files

At the broader community level, automated code review is mandatory. If the GitHub Codex review does not run automatically, contributors must run codex review --base origin/main locally and treat the results as a required review task.¹⁴

4. PR Patterns and Automation at Scale

OpenClaw handles roughly 14 new PRs per day, with 5,500+ open PRs and 19,000+ closed PRs at any given moment.¹⁵ That volume demands aggressive automation. The repository defines 58 labels in total, split by component (agents, cli, gateway, docker), channel (channel: telegram, channel: discord, channel: whatsapp-web), and type (docs, enhancement, bug).⁹

Several bots participate in the review process. The openclaw-barnacle bot handles auto-labeling and auto-response, Greptile-apps performs automated code analysis, and aisle-research-bot leaves review comments. The .github/workflows/auto-response.yml workflow automatically closes and locks issues matching specific patterns — TestFlight requests, third-party extension proposals, spam, and so on.¹⁵

VISION.md imposes strong limits: one PR should address exactly one issue or topic, unrelated fixes should not be bundled, and PRs changing roughly 5,000 lines or more are reviewed only in exceptional cases.¹⁶ The CI pipeline includes the main test suite, install smoke tests, workflow health checks, and the auto-labeler, and the coverage bar requires 70%+ across lines, branches, functions, and statements alike.¹²

5. Nine Best Practices Distilled from the Vibe Coding Workflow

OpenClaw’s official Vibe Coding skill (1,700+ stars on ClawHub) codifies this methodology into a set of rules.¹⁷ Combine that with Steinberger’s publicly documented workflow and the repository’s actual practices, and the most actionable patterns come out as follows.

5.1 Maintain a Living AGENTS.md File

This is the highest-leverage practice. Write it incrementally — add a rule every time an AI agent makes a mistake. Symlink it to CLAUDE.md for tool compatibility. Steinberger’s file runs about 300 lines and covers git conventions, test commands, architectural patterns, file length limits (roughly 500 LOC), and multi-agent coordination rules. It converts oral tradition into machine-readable institutional memory.¹²¹³

5.2 Use a Research → Plan → Implement Workflow

Have the AI explore the existing code before it implements anything. For example: “Read the auth module and explain how sessions work.” Then have it propose a plan: “List the files you will modify and what changes in each one.” Only after reviewing the plan do you move to implementation. “Catching a misunderstanding at the planning stage is 10x cheaper than debugging a cascade of errors.”¹⁷

5.3 Enforce Conventional Commits with Subsystem Scopes

The type(scope): description pattern — for instance fix(telegram): resolve TypeError in status command — produces a machine-parsable history that AI agents can understand and that enables automated changelog generation. Use a committer helper script so that staging stays limited to the intended files.¹²

5.4 Treat AI Agents Like Junior Engineers You Manage

This is Steinberger’s central analogy. Human effort should concentrate on system architecture and taste — the ability to distinguish a solution that works from a solution that is elegant. Delegate implementation, boilerplate, and refactoring to the agents.¹

5.5 Run Parallel Agents with Explicit Coordination Rules

Multiple agents can work concurrently even in the same folder — provided AGENTS.md spells out the following. Check git status and git diff before modifying anything. Make atomic commits. Do not touch stashes or switch branches. If you see a change you do not understand, treat it as another agent’s work and keep going. That is what turns “vibe coding” from a single-player game into a multiplayer orchestration problem.¹²³

5.6 Treat AI PRs as First-Class Citizens, with Transparency Guaranteed

Require contributors to disclose AI tool usage, state the level of testing, attach prompts or session logs, and confirm their understanding of the generated code. This is not about erecting barriers to entry. It gives reviewers the context they need to focus on the places where subtle AI-generated bugs are most likely to hide.¹⁴

5.7 Automate Everything That Does Not Require Taste

OpenClaw uses auto-labeling, auto-response workflows, automated code review bots, stale issue management, secret detection, dead code analysis, and duplicate checking. The human role is architecture and quality judgment. Everything else should be automated or delegated to agents.¹⁵

5.8 Know When to Intervene and When to Let It Flow

Hand scaffolding, UI components, boilerplate, and exploratory work to the AI. On the other hand, authentication, payments, data processing, database schemas, API permissions, and anything touching security require manual intervention. And always test after every change. AI produces code that is “seemingly perfect but subtly buggy.”¹⁷

5.9 Anchor Constraints in Your Prompts

Give explicit boundaries: line limits (“50 lines or fewer”), output format limits (“only the modified function, not the whole file”), scope anchors (“payment flow only, do not touch auth”), and style directives (“follow the existing patterns in UserService.ts”). Vague prompts produce vague results.¹⁷

Conclusion

The OpenClaw repository shows that production-scale vibe coding is not about abandoning engineering discipline but about relocating it. This project’s 1,200+ contributors, 19,000+ merged PRs, and 18,000+ commits run on exactly the principles that make large software projects possible: clear conventions, automated enforcement, structured communication, and explicit boundaries.

The innovation lies in who executes. AI agents handle implementation; humans take architecture, taste, and coordination. The AGENTS.md (CLAUDE.md) pattern — a living, incrementally grown instruction file that accumulates “organizational scar tissue”¹³ — is the most easily transferable practice of all.

Steinberger’s trajectory, moving from detailed specifications (June 2025) to short screenshot-driven prompts (late 2025) and then to running three to eight parallel agents with minimal supervision (2026), traces the learning curve every vibe coder will follow.³ The lesson is not that code no longer matters. It is that the center of gravity in software engineering has shifted: from mastery of writing code to mastery of designing the system that writes it. Accumulating a 300-line AGENTS.md, establishing a 70% coverage baseline, and designing multi-agent conflict rules is not bricklaying — it is architecture. And that architecture, too, demands a mastery that does not come easily.