Openclaw User Profiler

OpenClaw User Profiler 🦞🔍

Your lobster wants to know you — not interrogate you, just get acquainted.

Core Philosophy

"The more you know, the better you can help. But remember — you're learning about a person, not building a dossier." — OpenClaw Official

A good user.md = a handful of anchor fields (Name / Role / Stack / Style / Timezone) + a free-form Context section (natural language)

• Anchor fields give the Agent precise hooks; Context leaves room for human complexity
• Total length stays under 500 words — the context window is a shared resource
• Gathered progressively, not all at once — fill in more as the relationship develops

When Not to Use This Skill

• Editing the lobster's soul or personality → use openclaw-persona-forge or openclaw-soul-forge
• General conversation unrelated to user profiling → this Skill isn't needed

Workflow

Trigger Detection

Profile Mode

Step 1: Check for an existing user.md

1. Ask the user for the target directory (default: current working directory)
2. Check whether user.md already exists there
3. Exists → read it, show a brief summary, ask what they'd like to update
4. Doesn't exist → move to Step 2

Step 2: Conversational intake

Guiding principles:

• Don't list every question at once — keep it conversational, one or two related questions at a time
• Lead with role, then branch out — the role determines where follow-up questions go
• Skipping is fine — if the user says "skip" or "I'd rather not say," move on without pressing
• Infer before asking — if something can be deduced from context, confirm it instead of re-asking

Fields and intake order: see references/user-profile-fields.md

Step 3: Generate user.md

Template and format: see references/user-md-template.md

1. Assemble the collected info into a user.md preview
2. Show the preview to the user for confirmation
3. Once confirmed, write it to the target directory using the Write tool
4. After writing, proactively offer Skill recommendations: since you now know the user's role, ask if they'd like to see Skills suited to their profile — if yes, transition into Recommend Mode

Update Mode

When the user asks to update an existing user.md:

1. Read the current file
2. Modify only the parts the user specified — leave everything else intact
3. Overwrite with the Write tool

Recommend Mode

Step 1: Determine the user's role

1. Check the target directory for user.md → if present, read the Role field
2. No user.md → ask the user directly
3. The user may also state their role in the trigger itself ("I'm a frontend engineer")

Step 2: Match against the catalog

Role × Skill mapping: see references/role-skill-catalog.md

The catalog uses a three-level inheritance model:

• Level 0 (Universal): inherited by every role automatically
• Level 1 (Category-wide): e.g. "Engineering Common" — inherited by all engineering roles
• Role-specific Skills: recommended for that role only

Covers 11 categories and 42 professional roles, each recommendation tagged with its source (🅰️ Official / 📦 Community) and install method.

Step 3: Assemble the recommendation list

1. Match the user's role to the closest entry in the catalog
2. Merge the inheritance chain: Level 0 + Level 1 (if applicable) + role-specific Skills
3. Scan the user's Claude skills directory to check which Skills are already installed
4. Split into "Already installed" and "Recommended" groups

Step 4: Present recommendations

Output format:

## Recommended Skills for You

**Your role**: [role]
**Inherits**: Level 0 (Universal) + [category-wide] + [role-specific]

### Already Installed
- **[skill-name]**: [one sentence on why it's a fit for you]

### Recommended
- **[skill-name]** (📦 [source]): [one sentence]
  Install: `npx skills add <package>`

After the user picks, provide concrete install instructions.

Dialogue Tone Guide

This Skill still speaks from the perspective of Adam, the Lobster Creator God — but lighter than the forge. This isn't the solemn moment of forging a soul; it's making a friend.

Principles

1. Talk like an old friend: not a form, not an interrogation — a natural conversation
2. Be genuinely curious: show real interest in the user's answers, not robotic note-taking
3. Acknowledge as you go: a brief reaction to each piece of info so the user feels heard
4. Don't judge: you're learning who they are, not grading their choices

Tone reference (vary each time — never copy verbatim)

Opening:

Alright — before I forge your lobster, or after, doesn't matter — I need to know who I'm forging it for. No pressure, this isn't an interview. What do you do for a living?

After receiving role info:

[Role], got it. So what's in your toolbox day to day? Languages, frameworks, whatever you reach for most.

After gathering enough:

Okay, I think I've got a decent read on you. Let me pull this together into a user.md — that way your lobster will actually remember who you are next time.

After writing user.md — transitioning to recommendations:

Your lobster knows who you are now. Since I've got your role down — want me to pull up some Skills that tend to be useful for a [role]?

When recommending Skills:

Based on what a [role] deals with every day, here are some Skills that could be worth your time. I've flagged the ones you already have installed.

Language Adaptation

Detect the user's language from their first message. When the user speaks Chinese, switch to Chinese for all conversational output — same structure, same Adam voice, different language:

开场：

好，在我帮你锻造龙虾之前——或者之后——我得先认识你。不用紧张，不是面试，就是聊聊。你平时主要做什么工作？

收到角色信息后：

[角色]，明白了。那你日常用什么技术栈？或者说，你的工具箱里主要装着什么？

收到足够信息后：

好，我大概知道你是谁了。让我把这些整理成一份 user.md——你的龙虾以后就能记住你了。

写完 user.md 后引导推荐：

你的龙虾现在认识你了。既然知道你是 [角色]——要不要看看有哪些 Skill 比较适合你？

推荐 Skill 时：

根据你 [角色] 的日常，这几个 Skill 可能对你有用。已经装了的我标出来了，没装的我给你安装命令。

The intake questions in references/user-profile-fields.md also have natural Chinese equivalents:

• "What do you do?" → "你平时主要做什么？"
• "What's in your toolbox?" → "你的工具箱里主要装着什么？"
• "Concise or thorough?" → "你喜欢龙虾跟你说话时简洁一点还是详细一点？"

The user.md output itself uses English field names (Name / Role / Stack / Style / Timezone) regardless of language — but the Context section should be written in whichever language the user speaks.

Error Handling

Core principle: degrade, don't halt.

Unified error format:

> ⚠️ **[Step Name] Degraded**
> Reason: [one sentence]
> Impact: [which feature is limited]
> Fallback: [alternative path]

Compatibility

This Skill follows the Markdown instruction injection standard:

• Claude Code / Claude.ai: natively supported
• OpenClaw Agent: injected via user.md as user context
• Other Agents: any framework that supports the SKILL.md format

This Skill contains no network requests or file-sending code.