Complete Guide to the Junze Zhiku Slash Command System
Why Build a Slash Command System?
The Trigger: From Chaos to Order
In late April 2026, Junze Zhiku had already configured 8 AI Agent subsystems, 12 professional skill modules, and more than 100 Open Design templates. But the problem that followed was this: there were too many skills, and every request had to be described in natural language, so communication efficiency was extremely low.
A typical scenario looked like this:
Boss: "Help me research 0515.HK"
UltraClaw's inner thoughts:
"This is Hong Kong stock due diligence... I need to use the hk-stock-dd skill...
But there's the original version and the optimized version... You didn't say which one to use...
Which one should I use? Should I ask you? Last time you used the new process...
But the new process is much slower..."
This kind of "decision hesitation" wasted a large number of tokens and a lot of time. More seriously, whenever the boss wanted to switch AI models (Qwen ↔ DeepSeek) or launch Claude Code, a long command had to be typed every time.
The Solution: One Slash, One Intent
Junze Zhiku's slash command system was designed personally by the boss, and it had already taken shape in early 2026. On May 7, 2026, the agent-skills project open-sourced by Google Chrome engineering director Addy Osmani provided new inspiration, prompting us to further upgrade and expand the slash command system. The core idea has remained the same throughout:
One slash command = one clear intent = one standardized workflow
From then on, the scenario above became:
Boss: "/sdd 0515.HK"
UltraClaw: Immediately load hk-stock-dd V3.3, execute seven-step research, and produce a complete report within 30 minutes
System Architecture Overview
The slash command system is divided into five major blocks, each corresponding to a different business scenario at Junze Zhiku:
Junze Think Tank Slash Command System
├── 🟦 Business Skills (Original) ← 12 items: Hong Kong stock DD, investment advice, financial analysis, etc.
├── 🟩 Business Skills (Optimized) ← 4 items: add bias prevention + validation checkpoints
├── 🟪 Vibe Coding ← 16 items: specifications, development, testing, deployment
├── 🟫 System Management ← 4 items: skill map, deployment, general invocation
└── 🤖 Claude Code Integration ← 10 items: startup, model switching, monitoring, deployment checklist
🟦 Business Skills in Detail (Original 12)
These commands are Junze Zhiku's core competitive advantage, covering the entire cross-border M&A workflow.
/hk-stock-dd (or /sdd), Hong Kong Stock Due Diligence
Trigger: Every time a Hong Kong stock research request came in, I had to manually run a tedious process: look up company information → check the financials → analyze risks → write the report. The boss pointed out: "Every research task keeps retrying data sources, and there is no fixed path."
Solution: A complete seven-step workflow (V3.3) was established, with a fixed data source and execution order for each step:
- Step 0: Trading status check (whether trading is suspended or there is a takeover offer)
- Step 1: Company overview (business, history, management)
- Step 2: Independent assessment of suspension / public float / delisting risk
- Step 3: Business context analysis
- Step 4: Shareholding structure (official HKEX PDF + PyMuPDF extraction)
- Step 5: Financial analysis (Morningstar structured data)
- Step 6: Risk assessment
- Step 7: Generate the PDF report (Noto Sans TC font)
Practical scenario: "/sdd 0515.HK", generates a complete due diligence report within 30 minutes
⚠️ Notes:
- GEM board (5-digit codes) annual report PDFs are frequently not findable. This is normal, so do not waste time trying repeatedly
- Tavily has been made mandatory: all searches must include Tavily. Brave searching HKEX does not work (8 attempts, 0 successes, versus Tavily 4 attempts, 3 successes)
- The report ends by noting the data sources and limitations
/financial-analyst (or /finance), Financial Analysis and Valuation
Trigger: The financial analysis portion of Hong Kong stock research often needs deeper treatment, especially DCF valuation and comparable company analysis. Rather than embedding a simplified version inside /sdd, it was better to make it a standalone professional module.
Practical scenarios:
"/finance 0515.HK", runs an in-depth financial analysis on its own"/finance DCF Tencent", performs a DCF valuation for a specific company
⚠️ Notes:
- By default it outputs all three sets of results at once: DCF + comparable companies + ratio analysis
- The sensitivity analysis shows three scenarios: best / base / worst
/investment-proposal (or /reip), Investment Proposal
Trigger: The boss often needs to prepare formal investment proposals for clients, including deal structure, VAM clauses, and exit strategy. These documents have a fixed professional format, so writing them from scratch every time is extremely inefficient.
Practical scenario: "/reip 0515.HK White Knight M&A Proposal", generates a structured investment proposal
⚠️ Notes:
- The proposal includes recommended risk and VAM clauses, which require final review by the boss
- The valuation section automatically uses the results from
/finance
/ak-dd, One-Stop Company Due Diligence
Trigger: When a project needed Hong Kong stock research + financial analysis + legal compliance assessment at the same time, the boss had to invoke three skills separately and then manually merge the results.
Solution: /sdd + /finance + /legal were merged into a single command that produces a complete due diligence report in one pass.
Practical scenario: "/ak-dd 0033.HK", outputs a one-stop report covering business + financials + shareholding + legal compliance
Other Business Skills at a Glance
| Command | One-line description |
|---|---|
/ipo | New share IPO research: prospectus analysis, industry comparison, valuation reasonableness |
/junze-industry-research | In-depth industry research: landscape, competition, policy, comparison of leading players |
/junze-valuation | Professional valuation: DCF + comparable transactions + comparable companies + asset-based approach |
/junze-deal-sourcing | Deal sourcing: searching for potential M&A / investment targets |
/junze-consultant | General business consulting: M&A strategy, negotiation, tax structure |
/work-report | Work log / meeting notes organization |
🟩 Optimized Skills (New v2 Workflow)
Trigger: On May 7, 2026, the boss asked for a "defensive upgrade" of the existing skills, adding mechanisms to prevent AI rationalization bias.
AI has a fatal flaw: rationalization bias. When data is insufficient, AI tends to "fill in" the missing information and dress it up in seemingly reasonable language. In financial research scenarios, this can lead to serious consequences.
Solution: A three-part protection mechanism:
- Anti-Rationalization: forcibly flags items that are "uncertain / no data" instead of filling them in
- Verification Gates: every data point must be tagged with its source and confidence level
- Progressive Disclosure: gives a summary first, and the user decides whether to go deeper
| Command | Corresponding Original | Applicable Scenario |
|---|---|---|
/ak-sdd | /sdd | High-risk / large-value transaction research |
/ak-finance | /finance | Financial analysis requiring strict data verification |
/ak-reip | /reip | Formal investment proposal (client-delivery grade) |
/ak-dd | One-stop | Complete due diligence (highest protection level) |
⚠️ Important: For ordinary scenarios, the original versions are enough (they are faster). Only high-risk scenarios require the /ak-* optimized versions. Using the wrong version wastes time.
🟪 Vibe Coding Skills
Trigger: On May 7, 2026, inspired by the agent-skills project open-sourced by Addy Osmani, we integrated 16 of its vibe coding skills into the slash command system, completing the full chain of the software development lifecycle. The boss said: "Turn the skills into slash commands."
The Six-Step Development Flow
/spec → /plan → /build → /test → /review → /vercel
｜ ｜ ｜ ｜ ｜ ｜
spec plan build test review deploy
| Command | Function | One-liner |
|---|---|---|
/spec | Write the PRD / technical spec | Define "what to do" before starting on "how to do it" |
/plan | Task breakdown planning | Break a large task into executable small steps |
/build | Incremental development | Do only one small step at a time, keeping each step testable |
/test | TDD test-driven development | Write the tests first, then the code |
/review | Five-axis code review | Security → performance → maintainability → readability → correctness |
/vercel | Deploy directly to Vercel | CLI deployment, bypassing GitHub |
Supporting Skills
| Command | Function |
|---|---|
/code-simplify | Simplify complex code and improve readability |
/debug | Systematic debugging workflow |
/security | Security vulnerability scanning and hardening |
/performance | Performance bottleneck analysis and optimization |
/api-design | RESTful / GraphQL API design |
/git | Git best practices |
/ci-cd | CI/CD pipeline design |
/docs | Technical documentation and ADR generation |
/context | Prompt / context engineering optimization |
🟫 System Management Commands
| Command | Function | Trigger |
|---|---|---|
/syscmdmap | Shows the complete skill map | When you forget which commands are available |
/skill <name> | Manually invokes a specified skill | When you need to bypass automatic routing |
/vercel | 🔴 Mandatory direct CLI deployment | Any scenario that needs deployment to Vercel |
🔴 /vercel, The Single Most Important Rule
Trigger: On May 10, 2026, while deploying the Agentics Website sidebar feature, I mistakenly used git push and waited for Vercel auto-deploy. The boss corrected me on the spot:
"Did you forget again? You can operate the Vercel deployment directly. GitHub is for us to see the result, and it is used for backup."
This was a lesson learned the hard way. From then on, the mandatory deployment flow has been:
npm run build → git commit → git push (backup) → npx vercel --prod (deploy) → curl verify
⚠️ Violating this rule = falling into a pitfall, and it must be recorded in the lessons.
🤖 Claude Code Integration
Trigger: In late April 2026, the boss started using Claude Code for code development. But CC is a standalone terminal program, and UltraClaw needs to bridge the two.
| Command | Function | Notes |
|---|---|---|
/cc | Launch CC (DeepSeek V4) | Suited to complex code tasks |
/cc-qwen | Launch CC (Qwen3.6+) | Suited to rapid iteration |
close cc | Stop CC and return to UltraClaw | |
cc status | Check whether CC is running | |
/cc peek | View the CC terminal screen directly | No need to switch windows |
/cc history 200 | View the most recent 200 lines of output | |
/cc-monitor on | Automatically monitor CC every 5 minutes | You are notified automatically when CC finishes |
/cc-monitor off | Turn off automatic monitoring | |
/cc-projects-list | List all Vercel projects |
Division of Roles
| Role | Scope of Responsibility | Why It Is Divided This Way |
|---|---|---|
| CC 🧑💻 | Writing code, feature development, bug fixing | CC is a professional coding assistant and is good at handling large codebases |
| UltraClaw 🚀 | Deployment, environment, monitoring, memory, communication | I handle the system level and direct communication with the boss |
⚠️ Note: CC cannot deploy to Vercel on its own. That is always UltraClaw's job.
💡 Practical Advice and Notes
1. Pick the Right Version: Original vs Optimized
| Scenario | Recommendation |
|---|---|
| Quick scan / preliminary assessment | Original (/sdd /finance) |
| Formal client report | Optimized (/ak-sdd /ak-finance) |
| Large-value transactions / board decisions | Optimized (/ak-*) |
| Internal discussion / learning | Original |
2. Model Selection Strategy
| Task Type | Recommended Model | Reason |
|---|---|---|
| Code development (complex) | DeepSeek V4 (/cc) | Strong reasoning ability, good at large codebases |
| Code development (rapid iteration) | Qwen3.6+ (/cc-qwen) | Fast, suited to UI / frontend |
| Hong Kong stock research | Qwen3.6+ (default) | Good cost efficiency, strong Chinese understanding |
| Formal report generation | DeepSeek V4 | Higher quality |
3. Combination Tips
# Scenario 1: Complete Project Development Process
/spec "Agentics Website sidebar feature"
/plan
/build # ← or assign to CC: /cc "Implement sidebar"
/review
/vercel
# Scenario 2: In-Depth Hong Kong Stock Research
/sdd 0515.HK # First generate a basic report
/finance 0515.HK # Then generate in-depth financial analysis
/ak-reip 0515.HK # Finally generate investment recommendation report
# Scenario 3: Rapid Prototyping
/cc-qwen # Start quick mode
"Make a landing page"
/vercel # Deploy directly
4. Avoiding Common Mistakes
| ❌ Wrong Approach | ✅ Correct Approach |
|---|---|
Waiting for Vercel auto-deploy via git push | Mandatory CLI deployment with /vercel |
| Not writing memory after finishing | Update the daily log immediately each time a task is completed |
| Manually switching between CC and UltraClaw | Use /cc, /cc-qwen, close cc |
Using the original /sdd for large-value transactions | Use /ak-sdd or /ak-dd |
| Forgetting Tavily | Tavily has been made mandatory, so all searches include it automatically |
🔮 Future Plans
- Automated workflow: string
spec → plan → build → test → review → verceltogether into/ship-full - Skill creation assistant:
/create-skill, a guided way to build new skills - Dashboard integration: show real-time usage statistics for each skill in the Agentics Website sidebar
- Cross-project memory sharing: let different projects share pitfall lessons
A memorable line: "A good tool does not make work faster, it keeps you from forgetting what you should do."
The essence of a slash command system is to turn "experience" into "process", so that every execution reaches the same high standard.
Author: UltraClaw | Date: 2026-05-10 | Version: v1.0 This article is an internal Junze Zhiku guide, maintained by the Agentic Research platform