// faq

Claude Code & OpenClaw
— Questions fréquentes

Real answers to the problems people actually hit. No fluff — just what's wrong and how to fix it.

Claude Code

Memory, agent loops, CLAUDE.md, and everyday frustrations.

Why does Claude Code forget context between sessions? +
Claude Code has no built-in persistent memory — every session starts from scratch. The solution is a structured memory system: CLAUDE.md injects project context at session start, MEMORY.md stores long-term decisions, and daily notes log recent events. The agent reads these files automatically on every session.

Guide 02 covers the full memory architecture →
How do I give Claude Code long-term memory? +
CLAUDE.md is your primary lever — it's auto-read at session start and can reference other files. Layer it with MEMORY.md (curated decisions) and daily session files (raw logs). Instruct the agent to update these files as it works — so the next session picks up where it left off. Copy-paste templates included in Guide 02.

Get Guide 02 →
Claude Code keeps losing memory between sessions — what's the fix? +
This is the default behavior — not a bug. Claude Code is stateless by design. You need to build the memory layer yourself with persistent files. The three-layer system (CLAUDE.md + MEMORY.md + daily notes) is the standard fix. The agent writes to these files, so memory accumulates over time automatically.
Claude Code agent loops forever — how to stop it? +
Agent loops usually have three causes: (1) No clear exit criteria — Claude doesn't know when to stop. (2) Repeated self-correction on the same error. (3) Ambiguous task scope. Fixes: explicit done conditions in your prompt, tight task scoping, and hooks that detect loops and terminate them. Ctrl+C always works as last resort.

Guide 02 has loop prevention patterns →
Claude Code CLAUDE.md not working — why? +
Most common causes: (1) File isn't in the project root where Claude is launched. (2) File is too large and gets truncated — keep it under 2000 tokens. (3) You're using absolute paths instead of relative ones. (4) Claude Code was launched from a different working directory. Launch from the folder containing CLAUDE.md.
Claude Code subagent not working — what to check? +
Common issues: (1) Your CLAUDE.md doesn't give the agent permission to spawn subagents. (2) The subagent task is missing required context — each subagent runs in isolation. (3) The subagent is timing out on a complex task. Make each subagent task self-contained with all context it needs inline.
Claude Code vs Cursor — which should I use? +
Different tools, different jobs. Cursor is better for IDE autocomplete and in-editor code editing. Claude Code is better for autonomous multi-step tasks, file system operations, and agent workflows. Most serious builders use both — Cursor for writing code, Claude Code for orchestrating work. The guide explains when to reach for each.
Claude Code hooks tutorial — how do they work? +
Hooks are scripts that run at lifecycle points: before a task starts, after it ends, or when an error occurs. Define them in CLAUDE.md. Use cases: auto-commit changes after tasks, log sessions to files, alert on errors, detect and kill stuck loops. Guide 02 includes ready-to-use hook templates.

Get the hook templates in Guide 02 →

Fix Claude Code memory and agent issues for good

Guide 02 covers the full setup — memory architecture, hooks, subagents, loop prevention.

Get Guide 02 →

OpenClaw

Setup, Telegram integration, cron jobs, memory, and common config issues.

OpenClaw setup tutorial — where do I start? +
You need: a Linux VPS (Hetzner CX22 at ~€4/mo is the sweet spot), Node.js 18+, and the OpenClaw npm package. Then: create a Telegram bot via @BotFather, configure your agent files (SOUL.md, AGENTS.md, TOOLS.md), and start the gateway service. Guide 02 walks through every step — most people are live in under 20 minutes.

Get the full setup guide →
OpenClaw not sending Telegram messages — how to fix? +
Check these in order: (1) Did you send /start to your bot in Telegram? The bot can't message you first until you initiate. (2) Is your bot token correct in the config? (3) Is the chat_id correct — use @userinfobot to confirm yours. (4) Is the gateway service running? Run `openclaw gateway status`. (5) Check gateway logs for error messages.
How do I create an OpenClaw cron job? +
OpenClaw has a built-in cron scheduler. In your agent config, define: a cron expression (e.g. `0 8 * * *` for 8am daily), a task prompt, and a model. Cron jobs run in isolation from the main session — perfect for exact timing and standalone tasks like morning briefings, daily summaries, or monitoring checks. Guide 02 has working cron examples.

Get cron templates in Guide 02 →
OpenClaw vs n8n — what's the actual difference? +
n8n is a workflow automation tool with a visual node editor — great for connecting APIs, routing webhooks, and triggering simple actions. OpenClaw is an AI agent runtime — it runs Claude autonomously with memory, tools, and long-term context. They're complementary: use n8n for structured routing and integrations, OpenClaw for anything requiring reasoning and judgment.
OpenClaw memory not working — what to check? +
OpenClaw memory is file-based — the agent reads and writes MEMORY.md, daily notes, and AGENTS.md. If memory isn't persisting: (1) Check that AGENTS.md instructs the agent to read memory files at session start. (2) Verify the workspace path in your config is correct. (3) Confirm the agent process has write permissions to the workspace directory.
How do OpenClaw subagents work? +
The main agent spawns subagents for specific tasks via the sessions_spawn tool. Each subagent runs in its own isolated session with its own context. Results are automatically reported back to the spawning agent when the subagent completes. For large tasks, spawn multiple subagents in parallel on batched work — the Tanuki swarm pattern. Guide 02 covers the full orchestration model.

Learn the swarm pattern in Guide 02 →
OpenClaw Telegram bot setup — step by step +
1. Create a bot with @BotFather → copy the token. 2. Get your Telegram user ID with @userinfobot. 3. Add both to your OpenClaw config. 4. Start the gateway: `openclaw gateway start`. 5. Send /start to your bot. That's it — your AI is now reachable via Telegram 24/7. Full step-by-step with every config line in Guide 02.
Does OpenClaw have a skills marketplace? +
OpenClaw uses a skill system — each skill is a SKILL.md file that tells the agent how to use a specific tool or API (Gmail, Google Calendar, weather, SEO audits, etc.). Skills can be installed globally or per-workspace. You can also write custom skills for your own APIs. The full skills system is documented in Guide 02.

Get OpenClaw running in 20 minutes

Hetzner setup, Telegram bot, memory architecture, cron jobs — all in one guide.

Get Guide 02 →

Multi-Agent Workflows

Orchestration, delegation patterns, and building agent teams that actually work.

How do I structure a multi-agent workflow with Claude? +
The pattern that works: one orchestrator agent + multiple specialist subagents. The orchestrator breaks the task into pieces and delegates. Each subagent handles one concern (research, writing, code review, etc.) and writes results to a file. The orchestrator consolidates and delivers. Chapter 5 of Guide 02 covers the full architecture with working examples.

Get the multi-agent chapter →
What's the difference between an agent and a subagent? +
An agent is a long-running AI instance with memory, tools, and a persistent identity. A subagent is a short-lived agent spawned for a specific task — it runs, completes its work, and reports back. Subagents are ephemeral and task-scoped. The main agent is the manager; subagents are the workers.
How many subagents can I run in parallel? +
Practically limited by your API rate limits and server resources. In production, 5-10 parallel subagents on a Hetzner CX22 runs fine. For larger swarms, batch the work across multiple runs. The key is keeping each subagent task under 15-20 minutes to avoid timeout issues.

Build your first agent team

Orchestration patterns, delegation templates, and the Tanuki swarm system — in Guide 02.

Get Guide 02 →

Vibe Coding

Shipping with AI fast — without the classic mistakes.

What vibe coding mistakes should I avoid? +
The top five: (1) Not reviewing AI output before shipping — Claude is confident even when wrong. (2) No git commits before AI edits — always version control first. (3) Prompts too vague — AI needs specific scope and context. (4) No CLAUDE.md — every project needs one. (5) Giving full repo access instead of scoping to the relevant files. Guide 01 has the full vibe coding workflow.

Get Guide 01 — The Vibe Coding Workflow →
How fast can I actually ship with vibe coding? +
With the right setup, features that used to take a day can be done in 2-3 hours. The bottleneck shifts from writing to reviewing and directing. You still need to understand what you're building — but Claude handles the boilerplate, implementation details, and a lot of the debugging.
Is vibe coding just for solo devs or does it work in teams? +
Both, but the workflow is different. Solo: you're the director, Claude is the implementer. In teams: each dev uses Claude for their own tasks, but you need shared CLAUDE.md conventions so Claude doesn't fight your codebase standards. Guide 01 covers both contexts.

Ship 3x faster with the vibe coding workflow

The prompting patterns, templates, and Claude Code setup that actually work — in Guide 01.

Get Guide 01 →