OpenClaw Skill Development — Standard Operating Procedure

Follow this process for every OpenClaw skill development task. No shortcuts.

⛔ 铁律: 不可破坏 Memory

• 绝对不能删除、覆盖、截断 memory/ 目录下的任何文件和 MEMORY.md
• 只允许 append 操作，不允许 rewrite 或 truncate
• 部署 skill 时 rsync --delete 必须排除 memory/
• 迁移 workspace 时必须完整保留 memory/ 和 MEMORY.md

⛔ 铁律: 遇到问题先跑 openclaw doctor

• Skill 不加载、Gateway 不启动、部署失败 — 先运行 openclaw doctor
• doctor 会自动检测并修复常见问题，输出结果后再决定下一步

Architecture Overview

OpenClaw skills are modular capability packages defined by a SKILL.md file. They are loaded into an agent's system prompt to extend its capabilities.

Skill Resolution Order (highest precedence first)

1. Workspace skills: <agent-workspace>/skills/<skill-name>/SKILL.md
2. Managed skills: ~/.openclaw/skills/<skill-name>/SKILL.md
3. Bundled skills: <openclaw-install>/skills/<skill-name>/SKILL.md

Higher precedence overrides lower. Workspace skills are per-agent (configured via openclaw.json → agents.list[].workspace).

Key Files

File	Purpose
~/.openclaw/openclaw.json	Main config. Agent definitions, workspace paths, models
<workspace>/skills/*/SKILL.md	Skill definitions
<workspace>/SOUL.md	Agent personality
<workspace>/TOOLS.md	Tool usage guidance
<workspace>/AGENTS.md	Workflow and delegation rules
<workspace>/MEMORY.md	Persistent memory index
<workspace>/memory/*.md	Memory entries

Phase 1: Requirements

Before writing any code, gather these:

1. Target agent — which agent gets this skill? (check openclaw.json → agents.list)
2. Workspace path — where does that agent's workspace live? (check agents.list[].workspace)
3. Skill purpose — what capability does this add?
4. Trigger mode — always-on (metadata.clawdbot.always: true) or on-demand (user-invocable / model-triggered)?
5. Dependencies — does it require binaries (requires.bins), env vars (requires.env), or config (requires.config)?
6. Scope — agent-specific (workspace skill) or global (managed skill)?

Find agent workspace

jq '.agents.list[] | {id, workspace, model}' ~/.openclaw/openclaw.json

Phase 2: Design

SKILL.md Anatomy

skill-name/
├── SKILL.md          # Required — YAML frontmatter + markdown instructions
├── scripts/          # Optional — executable code
├── references/       # Optional — docs loaded on-demand
└── assets/           # Optional — files used in output

SKILL.md Format

---
name: skill-name
description: What this skill does and WHEN to use it. This is the primary trigger — be specific and comprehensive.
metadata: {"clawdbot":{"always":false,"emoji":"🔧","requires":{"bins":["jq"]}}}
user-invocable: true
---

# Skill Title

Instructions for the agent...

Frontmatter Fields

Field	Required	Description
name	Yes	Lowercase kebab-case, matches directory name
description	Yes	Primary trigger mechanism. Include WHAT it does AND WHEN to use it. The body is only loaded after triggering.
metadata	No	JSON string with OpenClaw-specific config (see below)
user-invocable	No	true = user can trigger via /skill-name command

Metadata Object (metadata.clawdbot)

Key	Type	Description
always	boolean	Auto-inject into every session's system prompt
emoji	string	Display emoji for the skill
skillKey	string	Config key in openclaw.json → skills.entries
primaryEnv	string	Primary env var (for API key gating)
os	string[]	OS filter: ["darwin"], ["linux"], etc.
requires.bins	string[]	ALL listed binaries must exist
requires.anyBins	string[]	At least ONE listed binary must exist
requires.env	string[]	Required environment variables
requires.config	string[]	Required config paths in openclaw.json
install	array	Auto-install specs (brew/node/go/uv/download)

Design Rules

1. Context window is a public good — skills share the prompt with everything else. Keep SKILL.md under 500 lines. Challenge every paragraph: "Does the agent really need this?"

2. Description is the trigger — the body only loads AFTER triggering. All "when to use" info goes in description, not the body.

3. Progressive disclosure — Level 1: metadata (~100 words, always loaded). Level 2: SKILL.md body (<5k words, loaded on trigger). Level 3: references/ (unlimited, loaded on demand).

4. Imperative voice — write instructions in imperative form ("Run X", "Check Y"), not descriptive ("This skill runs X").

5. No extraneous files — no README.md, CHANGELOG.md, INSTALLATION_GUIDE.md. Only SKILL.md and functional resources.

6. Zero hardcoding (Organization Iron Law #1) — skills must be portable. Never embed environment-specific values:

   Forbidden	Compliant alternative
   /Users/xxx/, /Volumes/EXT/	$HOME, task-param $OUTPUT_FILE
   user@gmail.com	$GOOGLE_DRIVE_ACCOUNT from openclaw.env
   127.0.0.1:8045	read from openclaw.json → models.providers
   sk-abc123... (API keys)	read from secrets/ dir, never in SKILL.md
   model-name-v3 (literal)	resolve via model-routing-governor slot
   PYTHONPATH=/abs/path	$OPENCLAW_PYTHONPATH from openclaw.env

   SSOT for runtime values: ~/.openclaw/openclaw.env (env vars) and ~/.openclaw/openclaw.json (config). Skills read from these; they never redefine them.

Phase 3: Implement

Step 1: Create skill directory

WORKSPACE=$(jq -r '.agents.list[] | select(.id=="<agent-id>") | .workspace' ~/.openclaw/openclaw.json)
WORKSPACE=$(eval echo "$WORKSPACE")  # expand ~
mkdir -p "$WORKSPACE/skills/<skill-name>"

Step 2: Write SKILL.md

Follow the format above. Key checks:

• name matches directory name exactly
• description is comprehensive (includes triggers/contexts)
• metadata is valid JSON (use echo '...' | jq . to verify)
• Body uses imperative voice
• Under 500 lines
• No "When to use" section in body (that belongs in description)
• References large content to references/ files

Step 3: Add resources (if needed)

• scripts/ — executable code that would be rewritten each time
• references/ — docs the agent reads on demand (keep SKILL.md lean)
• assets/ — templates, images, files used in output

Phase 4: Validate

Run these checks before considering the skill ready:

Structural validation

SKILL_DIR="$WORKSPACE/skills/<skill-name>"

# 1. SKILL.md exists
test -f "$SKILL_DIR/SKILL.md" && echo "OK: SKILL.md exists" || echo "FAIL: Missing SKILL.md"

# 2. Frontmatter has required fields
head -20 "$SKILL_DIR/SKILL.md" | grep -q "^name:" && echo "OK: has name" || echo "FAIL: missing name"
head -20 "$SKILL_DIR/SKILL.md" | grep -q "^description:" && echo "OK: has description" || echo "FAIL: missing description"

# 3. Name matches directory
DIRNAME=$(basename "$SKILL_DIR")
SKILLNAME=$(head -20 "$SKILL_DIR/SKILL.md" | grep "^name:" | sed 's/name: *//')
[ "$DIRNAME" = "$SKILLNAME" ] && echo "OK: name matches dir" || echo "FAIL: name '$SKILLNAME' != dir '$DIRNAME'"

# 4. Metadata JSON is valid (if present)
META=$(head -20 "$SKILL_DIR/SKILL.md" | grep "^metadata:" | sed 's/metadata: *//')
if [ -n "$META" ]; then
  echo "$META" | jq . > /dev/null 2>&1 && echo "OK: valid metadata JSON" || echo "FAIL: invalid metadata JSON"
fi

# 5. Line count check
LINES=$(wc -l < "$SKILL_DIR/SKILL.md")
[ "$LINES" -le 500 ] && echo "OK: $LINES lines (under 500)" || echo "WARN: $LINES lines (over 500, consider splitting)"

# 6. Hardcoding compliance check (Organization Iron Law #1)
echo "--- Hardcoding scan ---"
VIOLATIONS=$(grep -nE '/Users/[^$]|/Volumes/EXT|[a-zA-Z0-9._%+-]+@[a-zA-Z0-9.-]+\.[a-zA-Z]{2,}|127\.[0-9]+\.[0-9]+\.[0-9]+:[0-9]+|sk-[a-f0-9]{20,}|PYTHONPATH=/' "$SKILL_DIR/SKILL.md" 2>/dev/null)
if [ -z "$VIOLATIONS" ]; then
  echo "OK: no hardcoded paths/credentials/IPs found"
else
  echo "FAIL: hardcoded values detected — replace with env vars or task params:"
  echo "$VIOLATIONS"
fi

Dependency validation

# Check required binaries exist
META=$(head -20 "$SKILL_DIR/SKILL.md" | grep "^metadata:" | sed 's/metadata: *//')
if [ -n "$META" ]; then
  for bin in $(echo "$META" | jq -r '.clawdbot.requires.bins[]? // empty'); do
    which "$bin" > /dev/null 2>&1 && echo "OK: $bin found" || echo "FAIL: $bin not found"
  done
fi

Phase 5: Deploy & Verify

Deploy

The skill is already in the workspace directory. No copy needed. But the agent needs a new session to pick it up.

Restart gateway (if config changed)

# Only needed if openclaw.json was modified (e.g., workspace path changed)
pkill -TERM openclaw-gateway
sleep 3
# Gateway should auto-restart via launchd/systemd. If not:
# openclaw-gateway &

Trigger new session

Send /new to the agent via its messaging platform (Telegram, Discord, etc.). This creates a fresh session that re-resolves skills from the workspace.

Verify skills loaded

# Check session snapshot includes the new skill
AGENT_ID="<agent-id>"
cat ~/.openclaw/agents/$AGENT_ID/sessions/sessions.json | python3 -c "
import sys, json
data = json.load(sys.stdin)
session = data.get('agent:${AGENT_ID}:main', {})
prompt = session.get('skillsSnapshot', {}).get('prompt', '')
skill = '<skill-name>'
if skill in prompt:
    print(f'FOUND: {skill} loaded in session')
else:
    print(f'NOT FOUND: {skill} — check workspace path and gateway')
"

Verify skill content in session

# Check the latest session log for skill awareness
LATEST=$(ls -t ~/.openclaw/agents/$AGENT_ID/sessions/*.jsonl | head -1)
tail -5 "$LATEST" | python3 -c "
import sys, json
for line in sys.stdin:
    try:
        obj = json.loads(line.strip())
        role = obj.get('message', {}).get('role', '')
        if role == 'assistant':
            content = obj.get('message', {}).get('content', '')
            if isinstance(content, list):
                for c in content:
                    if c.get('type') == 'text':
                        print(c['text'][:300])
            elif isinstance(content, str):
                print(content[:300])
    except: pass
"

Common Pitfalls (from real experience)

Pitfall	Symptom	Fix
Skill in wrong directory	Not loaded in session	Verify agent's workspace in openclaw.json, put skill in <workspace>/skills/
Skill in shared workspace	Affects all agents	Use per-agent workspace via agents.list[].workspace
Gateway not restarted	Old session, old skills	pkill -TERM openclaw-gateway, wait for auto-restart
Session not refreshed	Skills in config but not in prompt	Send /new to agent to create fresh session
metadata not valid JSON	Skill loads but flags ignored	Test with echo '<metadata>' | jq .
Description too vague	Skill never triggers	Include specific trigger phrases and contexts
Body too long	Context bloat	Move details to references/ files, keep body < 500 lines
always: true on optional skill	Wastes context every session	Only use always for core behavioral skills
No dependency check	Skill fails at runtime	Use requires.bins / requires.anyBins in metadata
Memory files not preserved	Agent loses history during migration	Always copy memory/ and MEMORY.md when migrating workspaces
Hardcoded paths/emails/keys	Skill breaks on different machine or account	Use $HOME, $GOOGLE_DRIVE_ACCOUNT, $OPENCLAW_PYTHONPATH from openclaw.env; read model names from openclaw.json
API key in SKILL.md plaintext	Credential leak (CRITICAL)	Read from secrets/ dir via task param; never write into skill files
Lab path in production skill	Couples production to dev environment	Production skills (~/.openclaw/workspace-*/) must not reference /Volumes/EXT/openclaw-god/ or similar lab paths

Skill Categories & Templates

Category A: Behavioral Protocol (always-on)

Skills that define HOW the agent works. Examples: task execution protocol, communication standards.

• always: true — always in system prompt
• No user-invocable — not triggered manually
• Keep extremely concise — loaded every single session

Category B: Tool Integration (model-triggered)

Skills that teach the agent to use a specific tool. Examples: coding-agent, oracle, email.

• always: false — loaded only when relevant
• Comprehensive description with trigger phrases
• Include tool-specific examples and flags

Category C: User Command (user-invocable)

Skills triggered by user via slash command. Examples: /insights, /deploy.

• user-invocable: true
• Clear command syntax in body
• May also be model-triggered if description matches

Category D: Self-Improvement (hybrid)

Skills that analyze and optimize the agent. Examples: self-evolve, session analytics.

• Both always: true (for heartbeat mode) and user-invocable: true (for deep mode)
• Reads session logs at ~/.openclaw/agents/<id>/sessions/*.jsonl
• Tracks state in memory/ directory

Additional Resources

This skill now includes comprehensive supporting resources for progressive disclosure:

Utility Scripts (scripts/)

Validation and deployment utilities (all support local and remote agents):

• scripts/validate-skill.sh - Validate skill structure and metadata

  # Local validation
  ./scripts/validate-skill.sh /path/to/skill-name
  
  # Validate in agent workspace
  ./scripts/validate-skill.sh -a momiji skill-name
  
  # Validate on remote agent
  ./scripts/validate-skill.sh -r user@remote-host -a momiji skill-name
  

• scripts/deploy-skill.sh - Deploy skill to local or remote agent

  # Deploy locally
  ./scripts/deploy-skill.sh ./my-skill momiji
  
  # Deploy to remote agent (rsync)
  ./scripts/deploy-skill.sh -r user@remote-host ./my-skill momiji
  
  # Deploy to remote agent (git)
  ./scripts/deploy-skill.sh -m git -r user@remote-host my-skill momiji
  

• scripts/verify-skill-loaded.sh - Verify skill loaded in agent session

  # Verify local agent
  ./scripts/verify-skill-loaded.sh -a momiji skill-name
  
  # Verify remote agent
  ./scripts/verify-skill-loaded.sh -r user@remote-host -a momiji skill-name
  

Key features:

• ✅ No hardcoded paths - all paths resolved dynamically
• ✅ Support for remote OpenClaw agents via SSH
• ✅ 3 deployment methods: rsync/scp/git
• ✅ Comprehensive validation and verification

Reference Documentation (references/)

Detailed guides for progressive disclosure (load when needed):

• references/advanced-patterns.md - Advanced skill development patterns

  • Progressive disclosure strategies
  • Advanced metadata configuration
  • Remote deployment patterns
  • Multi-agent skill distribution
  • Security and performance optimization

• references/troubleshooting-guide.md - Common issues and solutions

  • Skill not loading issues
  • Deployment failures
  • Validation errors
  • Remote agent issues
  • Performance problems
  • Comprehensive diagnostic procedures

• references/real-world-examples.md - Production skill examples

  • Complete working examples from production
  • Deployment workflows (local/remote/multi-agent)
  • Pattern explanations and use cases

Working Examples (examples/)

Complete, copy-paste skill templates:

• examples/behavioral-protocol.md - Category A: Always-on behavioral skill

  • Example: task-execution-protocol
  • Ultra-concise (<300 lines)
  • Imperative rules

• examples/tool-integration.md - Category B: Tool integration skill

  • Example: api-client-integration
  • Comprehensive trigger phrases
  • Tool-specific commands

• examples/user-command.md - Category C: User-invocable command

  • Example: deploy command
  • Interactive workflow
  • Error handling

• examples/self-improvement.md - Category D: Self-improvement skill

  • Example: self-evolve
  • Dual mode (heartbeat + deep)
  • Session log analysis

Quick Reference

For comprehensive quick reference including:

• Minimal viable SKILL.md templates
• Common metadata patterns
• Validation checklists
• Deployment commands (local/remote/multi-agent)
• Troubleshooting one-liners
• Component organization guidelines
• Script usage examples

Read: references/quick-reference.md

---

For detailed patterns and techniques, consult the reference files and examples above.

Skill 清单查询

需要查看当前已安装的所有 skill（workspace 级 + managed 级 + bundled 级）时，读取 references/list-skills-runbook.md。