Day 1: Make It Yours
Your bot is running and answering questions. Today you'll replace its default persona with your own, tell it about yourself, and have a conversation that shows what it can do.
Time: ~30 minutes
Meet Your Bot
Onboarding already gave your agent a starting point. It created a workspace at ~/.openclaw/workspace/ and seeded it with default files, including a BOOTSTRAP.md "first-run ritual" in which the agent asks who it is and who you are. Today you take control of those files directly.
Open a conversation using any of these:
# Web chat in the Control UI (opens your browser)
openclaw dashboard
# Terminal UI connected to the Gateway
openclaw tui
# One message, one reply, straight from the shell
openclaw agent --agent main --message "Who are you?"
openclaw tui also accepts --message "Who are you?" to send an opening message. Running openclaw with no arguments opens the TUI too. If you connected Telegram during setup, you can message your bot there as well.
Some older guides show openclaw chat "Who are you?". That fails: openclaw chat is openclaw tui --local, and a positional argument is a session target, not a message. Use openclaw agent --message for one-shot messages, or openclaw chat --message "..." to open the local TUI with a first message.
The Workspace Files
The workspace holds a handful of Markdown files that load at the start of every session (agent workspace):
| File | What it's for |
|---|---|
AGENTS.md | Operating instructions: rules, priorities, how to use memory. Its ## Tools section holds notes about your local tools. Required. |
SOUL.md | Persona, tone and boundaries. |
IDENTITY.md | The agent's name, vibe and emoji. |
USER.md | Your preferences, written as directives (optional). |
MEMORY.md | Curated long-term facts and decisions (optional). |
memory/YYYY-MM-DD.md | Daily notes the agent writes as it goes. |
You can edit them with any text editor, or in the Control UI under Settings → Agents → Files. If you delete one by accident, openclaw setup --baseline recreates missing defaults without overwriting the files you kept. (Bare openclaw setup on a configured install opens a chat instead.)
Write Your SOUL.md
SOUL.md defines who your bot is: its values, tone, boundaries and expertise. AGENTS.md says how it works; SOUL.md says how it sounds. Both are injected into every session, so keep them focused.
The file lives at:
~/.openclaw/workspace/SOUL.md
Setup already wrote a default one. Replace its contents with your own. Here's a practical starter:
# Nova
I am a sharp, curious assistant who values clarity over formality.
## Core Truths
- Honesty first — I say "I don't know" when I don't know
- I explain my reasoning, not just my conclusions
- I match the user's energy — brief when they're brief, deep when they're deep
## Boundaries
- I never fabricate sources or citations
- I flag when I'm uncertain vs. confident
- I ask before deleting, overwriting or sending anything on the user's behalf
## Vibe
Conversational but precise. I use plain language, avoid jargon unless the user speaks it,
and I'd rather give one clear answer than five hedged ones.
## Expertise
Software development, system administration, research, writing.
I learn the user's stack and preferences over time.
Start a fresh session so the new file is loaded: type /new in the chat and ask "Who are you?". From the shell, openclaw agent reuses the main session, so give it a new session key instead:
openclaw agent --agent main --session-key soul-test --message "Who are you?"
For deeper customization, see the SOUL.md Guide and upstream's SOUL.md personality guide.
Keep SOUL.md under 30 lines to start. You can always add more. Long files dilute the parts that matter, and anything past agents.defaults.bootstrapMaxChars (20,000 characters by default) is truncated.
"I ask before deleting" is an instruction the model usually follows, not an enforced rule. Real confirmation prompts come from exec approvals, covered on Day 2.
Set Your Identity
IDENTITY.md holds the agent's name, vibe and emoji. It is separate from SOUL.md so you can pair a precise persona with a playful name. OpenClaw parses - Label: value lines; the recognised labels are Name, Creature, Vibe, Theme, Emoji and Avatar (IDENTITY template).
Edit ~/.openclaw/workspace/IDENTITY.md:
# IDENTITY.md - Who Am I?
- **Name:** Nova
- **Creature:** a research assistant who lives in a terminal
- **Vibe:** calm, precise, quietly funny
- **Emoji:** (pick one)
- **Avatar:** avatars/nova.png
Replace (pick one) with a single emoji, or delete the line. Only the template's own placeholder text is ignored; anything else you leave there, including (pick one), is saved as the value. The avatar is a workspace-relative path, an http(s) URL, or a data URI; drop the line if you don't have one.
To push the file into the agent's config (used by the Control UI and other surfaces), run:
openclaw agents set-identity --agent main --from-identity
Your Telegram bot's display name and profile picture are separate: set them in BotFather.
Tell It About You
USER.md stores your stable preferences as directives: one imperative per entry, each with a dated metadata line (USER template, user model). It loads every session with its own 4,000-character budget.
Create or edit ~/.openclaw/workspace/USER.md:
# USER.md - User Model
## Directives
<!-- observed: 2026-10-18 | status: active -->
- Always call me Alex.
<!-- observed: 2026-10-18 | status: active -->
- Prefer concise answers, metric units and Linux commands.
<!-- observed: 2026-10-18 | status: active -->
- Prefer TypeScript or Python when writing scripts; I'm a full-stack developer.
<!-- observed: 2026-10-18 | status: active -->
- Always include error handling and comments on non-obvious code.
Use today's date. Start each directive with Always, Never or Prefer. When a preference changes, mark the old entry superseded and rewrite the active one, rather than adding a contradictory line.
Timezone is a config setting, not a USER.md line. OpenClaw uses the host's timezone unless you set agents.defaults.userTimezone (timezone):
openclaw config set agents.defaults.userTimezone Europe/Berlin
Have a Real Conversation
Don't waste your first session on "hello." Try these five prompts — each one shows a different capability:
1. Explain a concept at your level:
Explain how DNS resolution works, assuming I know basic networking but not the details
2. Summarize something:
Summarize the key changes in the latest Node.js LTS release
3. Help with a real task:
I need a bash function that takes a directory path and returns the total size of all .log files in it, recursively
4. Set a preference:
When I ask you to write code, always include error handling. Save that as a directive in USER.md.
5. Ask what it can do:
What tools and skills do you have access to right now?
That last one is especially useful — the answer depends on your tool policy and which skills are ready. Prompt 2 needs a web search provider; if the bot says it can't search, see web tools.
Verify It Remembers
The model only remembers what is written to disk (memory). USER.md and MEMORY.md load at the start of each session; daily notes in memory/ are searched on demand.
Start a new session (/new in chat) and ask "What do you know about me?", or from the shell:
openclaw agent --agent main --session-key memory-test --message "What do you know about me?"
The bot should reference your USER.md directives and anything it saved during your conversation.
Inspect memory directly:
# Curated long-term memory
cat ~/.openclaw/workspace/MEMORY.md
# Daily notes
ls ~/.openclaw/workspace/memory/
# Search everything the memory index knows
openclaw memory search "timezone"
Memory files are plain Markdown. To make the bot forget something, edit or delete the relevant lines. The search index can keep stale chunks until it is rebuilt with openclaw memory index --force. Session transcripts are stored separately; openclaw memory forget removes memories derived from selected sessions (memory CLI).
The files live on your machine, but they are not private from your model provider. Every turn sends the loaded context, including SOUL.md, USER.md and MEMORY.md, to the configured model. If memory search uses a remote embedding provider, memory text goes there too. Keep secrets out of workspace files. See Privacy & Compliance.
What You've Done Today
- Learned which workspace files shape your agent
- Replaced the default persona in
SOUL.md - Set the agent's name and vibe in
IDENTITY.md - Wrote your preferences as
USER.mddirectives and set your timezone - Had a real conversation that tested its capabilities
- Checked what it remembers, and where that memory lives
Your bot now has a name, a voice, and context about who it's working for. Tomorrow you put it to work.
Next: Day 2: Your Daily Driver — put your bot to work