A practical guide to working with an AI assistant
A short, honest field-guide to using a modern LLM agent well — whether it's running on OpenClaw, Hermes, Claude Desktop, or something custom. Read top-to-bottom in ~15 min, or jump into a specific page from the nav.
Who this is for
Anyone who's started using an AI assistant — or is about to — and wants to get real leverage from it instead of treating it like a fancy search box. No prior AI background required.
#Why this guide exists
Modern AI assistants are powerful but easy to misuse. People ask too much in one shot, too vague, with too little context — then conclude "AI is bad at this" when really the model just couldn't see what was in the user's head.
The flip side: people also under-use them. They use an AI like a search engine when it can run a 5-step process autonomously — manage an inbox, draft and send emails, watch over projects, ping when something needs attention, generate reports nightly.
This guide covers both edges. What an AI assistant is good at, what it's bad at, how to prompt it well, how its memory actually works, and the small habits that compound into a 10× better working relationship.
#What runs under the hood
The advice here is platform-agnostic, but the worked examples and setup instructions assume OpenClaw — an open-source AI gateway that routes between messaging channels (WhatsApp, Signal, Telegram) and LLM providers (Anthropic, OpenAI, Gemini, local models).
Alternatives that follow the same patterns and where this guide still applies:
- Hermes — another open-source agent runtime, more focused on cloud + multi-user.
- Claude Desktop + MCP — Anthropic\'s native app with tool support.
- OpenWebUI / LibreChat — self-hosted chat UIs with plugin systems.
- Cline / Cursor / Aider — coding-focused agents.
- Custom builds on top of the Anthropic / OpenAI / Gemini SDKs.
The vocabulary (memory, prompting, tool calls, agents, hallucination) is the same across all of them. Pick whichever runtime fits your situation; everything in this guide still applies.
#The map
- Setup & install — how to get started from zero.
- What an AI assistant is good at — the strengths to lean on.
- What it's bad at — failure modes to plan around.
- How to prompt it — practical patterns that work and ones that don't.
- How memory works — what it remembers, what it forgets, where it's stored.
- Privacy & security — the threats and how to defend against them.
- Best practices — the habits that compound.
- Under the hood — what OpenClaw is, what makes it tick.
- Glossary — AI/agent terms worth knowing.
- Conclusion — the three things worth remembering.
#Six things worth knowing on day one
- Be specific or be surprised. "Write me an email" gets you a generic email. "Email the hotel asking for a 20:00 dinner reservation on Friday for 4 people, polite tone, in Portuguese" gets what you want.
- The assistant isn't always right. Push back. It sounds confident even when guessing. If something feels off, ask it to verify.
- Long messages drop on some channels. WhatsApp in particular chokes on multi-section walls. Keep replies short; put long content on pages.
- Each session starts fresh. The assistant's "long-term memory" is files on disk it loads on startup. If something matters, ask it to write it down.
- It can act, not just answer. Send emails, create calendar events, kick off builds, set reminders, manage projects. Don't use it like a search engine.
- External actions need approval. Anything that leaves the machine — email send, public post, calendar invite to someone — the assistant should check first. By design.
One frame to hold the rest of the guide: an AI assistant is a junior colleague with infinite patience and a goldfish memory. Brief it well at the start of each conversation, give it access to the right tools, and it'll do most things well. Skip the brief and it'll guess — sometimes well, sometimes embarrassingly.