From a Single Agent Session to a Governable Skill Registry: The SkillClaw × Nacos Automatic Evolution Loop

Problem 1 – Generating Skills

Real‑world agent tasks produce tacit knowledge (checklists, tool pre‑conditions, output formats, failure cases) that is not captured in design documents. Without a systematic mechanism this knowledge remains in a person’s memory.

SkillClaw is an open‑source framework (paper "SkillClaw: Let Skills Evolve Collectively with Agentic Evolver") that acts as a local proxy intercepting an agent’s model calls. It records each conversation turn, tool invocation, result, error, Skill hit, modifications, and scoring information. The Evolve Server analyses the recorded data and decides whether to create a new reusable Skill, improve an existing one, or discard the candidate.

Discovers reusable patterns from real agent usage and turns them into Skills.

Continuously improves existing Skills instead of rewriting prompts.

Provides a unified entry for Skill generation and synchronization across multiple agent clients.

Problem 2 – Sharing Skills

Even if a Skill is generated, a personal file cannot be safely reused by other team members. Nacos AI Registry treats a Skill as an AI Resource that can be registered, versioned, labeled, and moved through lifecycle states (draft → reviewing → online). It records who created, reviewed, and published each Skill, enabling audit and rollback.

Label‑based retrieval (e.g., latest) and version management.

Audit trails and traceability of creation and publishing decisions.

Pipeline plugins for security, compliance, and quality checks before a Skill is released.

Unified registry where any team member can discover, pull, and reuse Skills.

Closed‑Loop Integration

SkillClaw extracts candidate Skills from agent sessions.

The candidate is submitted to Nacos, entering the draft → review → online workflow.

Agents only read Skills; they cannot publish or delete them.

Governance ensures that only vetted Skills reach production, while continuous feedback can improve them.

The result is a pipeline that turns personal agent experience into a team‑wide, governed AI asset.

QuickStart Demo – From a Weekly‑Report Conversation to a Reusable Skill

Step 1: Install and start Nacos

curl -fsSL https://nacos.io/nacos-installer.sh | bash

Note the ServerPort, Console URL, username, and password displayed after startup.

Step 2: Clone and configure SkillClaw

# Clone the repository
git clone [email protected]:AMAP-ML/SkillClaw.git
cd SkillClaw
python3 -m venv .venv
source .venv/bin/activate
pip install -e .

# Set environment variables (replace placeholders)
export SKILLCLAW_HOME=$(pwd)
export SKILLCLAW_API_URL="YOUR_API_URL"
export SKILLCLAW_API_KEY="YOUR_API_KEY"
export SKILLCLAW_MODEL_ID="YOUR_MODEL_NAME"
export SKILLCLAW_DEMO_NACOS_USER_NAME="YOUR_NACOS_USER_NAME"
export SKILLCLAW_DEMO_NACOS_PASSWORD="YOUR_NACOS_PASSWORD"
export SKILLCLAW_DEMO_NACOS_SERVER_PORT="YOUR_NACOS_PORT"

Step 3: Configure and start the SkillClaw proxy

mkdir -p /tmp/skillclaw-demo/{skills,records,share,learn,verify}
skillclaw config llm.api_base "$SKILLCLAW_API_URL"
skillclaw config llm.model_id "$SKILLCLAW_MODEL_ID"
skillclaw config llm.api_key "$SKILLCLAW_API_KEY"
skillclaw config claw_type Claude

# Nacos registry configuration
skillclaw config sharing.enabled true
skillclaw config sharing.skill_backend nacos
skillclaw config sharing.nacos_server http://localhost:$SKILLCLAW_DEMO_NACOS_SERVER_PORT/nacos
skillclaw config sharing.nacos_username "$SKILLCLAW_DEMO_NACOS_USER_NAME"
skillclaw config sharing.nacos_password "$SKILLCLAW_DEMO_NACOS_PASSWORD"
skillclaw config sharing.nacos_publish_mode review
skillclaw config sharing.nacos_label latest

# Session and sync settings
skillclaw config sharing.session_upload_interval 3
skillclaw config sharing.skill_reload_mode poll
skillclaw config sharing.skill_reload_interval_seconds 60
skillclaw config sharing.auto_pull_on_start true

# Path settings
skillclaw config skills.dir /tmp/skillclaw-demo/skills
skillclaw config sharing.local_root /tmp/skillclaw-demo/share
skillclaw config record.dir /tmp/skillclaw-demo/records
skillclaw config evolve.server_url http://127.0.0.1:32218

# Start the proxy
skillclaw start --daemon --port 32217 --log-file /tmp/skillclaw-demo/proxy.log
skillclaw status

Step 4: Start the Evolver server

skillclaw-evolve-server \
  --use-skillclaw-config \
  --port 32218 \
  --interval 60 \
  --verbose > /tmp/skillclaw-demo/evolve.log 2>&1

Step 5: Conduct a three‑turn conversation in Claude Code to define a weekly‑report format

帮我写本周工作周报到weekly-report.md。飞书格式。#weekly-report
做了：1.完成user-auth-service OAuth2接入(UAT环境)
2.修复JIRA-2847登录超时、JIRA-2851密码重置、JIRA-2860验证码冷却
3.参加PR-1247和PR-1253 review。
格式：标题=[姓名]W{周数}周报；四个section：本周完成/进行中/下周计划/风险与协调；
每条=[代号]描述|关联JIRA|完成度%；结尾写---周报已同步至飞书#weekly-report

Claude returns a correctly formatted markdown file. After the third turn SkillClaw uploads the session, the Evolver generates a Skill named weekly-report-md, and the Skill is automatically submitted to Nacos in draft state.

Step 6: Verify the Skill in Nacos

Log into the Nacos console, navigate to “Skill Management”, and see the newly created Skill with status “已通过”. Publish it; the version becomes 0.0.1 and the label latest is attached.

Step 7: Reuse the Skill in a new Claude session

写本周周报到weekly-report.md，做了：1.完成OAuth2接入 2.修复JIRA-2847/2851/2860 3.review PR-1247/1253

Claude instantly recognizes the “weekly‑report” scenario, pulls the published Skill, and produces a compliant report with a single‑sentence prompt, reducing the prompt length from >200 words to 1 sentence.

Broader Use Cases

Onboarding new team members : Skills capture senior engineers’ tacit knowledge (testing steps, configuration order) and make it instantly reusable.

Repeated incident handling : Troubleshooting patterns become Skills, enabling agents to follow proven diagnostic paths.

Platform‑to‑business propagation : SDK usage, configuration standards, and release checks are distilled into Skills that Nacos distributes across business teams.

Unified agent behavior : Different front‑ends (Claude Code, Cursor) converge on the same Skill set, ensuring consistent output.

Support knowledge base : Frequently asked support questions are turned into Skills, audited, and shared team‑wide.

Governance and Safety

Nacos provides a controlled lifecycle for Skills: draft → reviewing → online, with versioning, labeling, audit trails, and rollback capability. Pipeline plugins (e.g., skill‑scanner) can enforce security rules, compliance checks, and quality scans before a Skill is released.

Future Directions

Link each Skill version to its originating session, resources used, and evaluation score, enabling full traceability of how a Skill was generated and why it was published.

Extend the evolution loop from individual Skills to full AgentSpec configurations (model settings, Skill collections, prompts, security policies), allowing the entire agent definition to evolve from real usage.

Conclusion

SkillClaw extracts experience from agent sessions; Nacos supplies governance, versioning, and distribution for team‑wide adoption. Together they form a closed‑loop pipeline that transforms personal agent knowledge into shared, evolvable AI assets.

References

SkillClaw paper: https://arxiv.org/abs/2604.08377 SkillClaw source code: https://github.com/AMAP-ML/SkillClaw Nacos AI Registry:

https://github.com/alibaba/nacos