Claude In Action Sai Gon #01 — Offline Workshop Recap (EN)
Recap of the first offline workshop: 3 practitioners share real-world Claude workflows for running agencies, SOP-as-Skill, and saving hundreds of millions VND/year.
Speakers
- Tony — CEO (Chief-of-Everything) @ Vibery Product Studio. Host & organizer.
- Duy Nguyen — Co-founder/CTO @ TOPGROUP, XINCHAO, DIGITOP. Build In Public VN (65K+ members). Building GoClaw.
- Chinh Dang — CEO @ SYNITY. CRM/SOP specialist. Live demo only (sensitive client data, no slides shared).
TL;DR
Three practitioners shared how they use Claude (Code, Cowork, Chat) to run their agencies: automating 80% of agency workflows with SOP-as-Skill, saving hundreds of millions VND/year on server costs, generating contracts in 5 minutes instead of an hour, and onboarding new hires with skills instead of training manuals. Core message: AI is not a tool — it’s a partner. Treat it like a capable intern, ask open-ended questions instead of giving orders, and package your expertise into SOPs for AI to execute while humans judge.
10 Takeaways
1. “Prompt Engineering is dead. Context Engineering is king.”
A year ago, everyone focused on crafting perfect prompts. Now models are powerful enough that a rough prompt works fine. The game has shifted to Context Engineering: managing context, providing the right SOPs, data sources, and reference files. A 1-million-token context window (~7 books) lets you feed sufficient context instead of writing the perfect prompt.
2. “Skills = SOPs written in natural language. Nothing fancy.”
Anthropic’s “Agent Skill” sounds impressive, but it’s really just bullet points describing a process — step 1 does what, step 2 does what, if-else logic. Three-part structure: Name + Description (when to trigger) + Body (instructions). Only ~10% of attendees had used Agent Skills. “Progressive Disclosure” framework: Skills don’t read every file — they read a table of contents first, then only load the reference files needed, saving tokens.
3. The “Hammer and Nail” Paradox: Fewer tools, more power
Tony uses one tool — Claude Code — for everything: code, content, contracts, server management, DNS, CRM, security audits. “When all you have is a hammer, everything is a nail.” Instead of learning 10 tools, master one and reshape every problem to fit. Ratio: 80% of agency work can be automated, 20% (closing deals) still needs humans.
4. “SOPs are no longer written for people to read — they’re written for AI to execute.”
Paradigm shift: SOPs used to be for training staff. Now they’re for AI to follow. When onboarding new hires, you don’t sit beside them step by step — hand them the skill, they interact with AI, AI executes the SOP. Chinh Dang shared: 4-5 years of CRM deployment experience lived in people’s heads, never documented — AI helped extract it into structured SOPs.
5. “Don’t lead with solutions — ask open-ended questions”
When you have deep domain expertise, you think you know the answer → tell AI → AI says “you’re absolutely right” → you learn nothing new. The right way: Pretend you don’t know, ask “how would you solve this?” → AI surfaces options you hadn’t considered. The human’s job = judgment (evaluating right/wrong, good/bad). AI executes, humans judge.
6. “Paying $100/month for an intern who never talks back — best deal in history.”
AI is like a sweet intern: fresh out of school, thinks they know everything, but always says “you’re absolutely right.” Cost ratio: ~$100/month (Claude subscription) vs $400-800/month for an actual intern in Vietnam. But this AI intern works 24/7, no holidays, no drama. Warning: AI (especially Opus 4.6) hallucinates — says “I’ve finished” then finds more problems to “fix.”
7. Plesk License dies, Telegram bot lives
Tony cut Plesk licenses ($60/month/server) after realizing Claude Code could replace it. Built a Telegram bot for server management in 2 hours: add domains, fix issues, troubleshoot — all from Telegram commands. Numbers: Saved hundreds of millions VND/year from cleanup, consolidating ~200 WordPress sites.
8. Identity Shift: “From AI-last to AI-first in one week on bed rest”
Chinh Dang — tech and automation background — was still “AI-last” before Tet: only using AI when problems exceeded his abilities. After a week of back injury confined to bed, forced to use AI for everything → complete mindset shift to AI-first. “Cooking rice” framework: If you keep turning the stove on for 5 minutes then off, rice never cooks. You have to use AI continuously, immerse yourself, to become proficient.
9. “Cost-saving AI for teams = GitHub + Telegram + 1 machine running 24/7”
Instead of buying Claude at $100/person × 10 people = $1,000/month, Tony built a system: employees send commands via Telegram bot → creates issues on GitHub → 1 machine running Claude Code Max ($200/month) pulls tasks from GitHub → executes per SOP → returns results. Cost: ~$200-250/month for the entire team.
10. “Read other people’s skills before using them — learn and survive”
Duy Nguyen warned about 2 problems when downloading skills from the internet: (1) Security: Many skills contain malicious code that can steal API keys or crypto wallets. (2) Slop: AI-generated skills often look impressive initially but produce poor real-world output — you must test and iterate. Always read skill file contents before using them.
Quotable Quotes
“When all you have is a hammer, everything is a nail.” — Tony
“Paying about $100 a month for an intern who always says ‘you’re absolutely right’ — that’s a pretty good deal.” — Tony
“AI can’t replace me for drinking sessions and building relationships. AI can’t go to court for me.” — Tony
“Everything works fine until demo time.” — Tony
“Before AI, I worked 8 hours a day. After adopting AI, I saved so much time… so now I work 16 hours a day.” — Chinh Dang
“Once you go down this path, there’s no coming back.” — Chinh Dang
“Most AI-generated skills are slop. They look impressive at first but the actual results tell a different story.” — Duy Nguyen
“Don’t outsource the thinking part. A brain that isn’t used gets weaker.” — Duy Nguyen
“In 2024, AI was a tool. In 2026, AI is a partner.” — Tony
“My job is to judge. That’s the one thing AI hasn’t replaced yet.” — Tony
Key Numbers
| Metric | Number |
|---|---|
| Event registrations | 50+ people (cap 30, overflow continued) |
| Attendees who’d used Agent Skills | ~10% |
| Claude cost/month (team setup) | ~$200-250 USD |
| Contract generation: before vs after AI | 30-60 min → 5 min (15 sec ideal) |
| Agency workflow automation | 80% |
| Annual savings from server cleanup | Hundreds of millions VND |
| WordPress sites migrated | ~200 sites |
| CPU usage after security fix | 78% → 0.4% |
| Server breach investigation time | 1-2 days → 5-10 min |
| Time to build Telegram server bot | 2 hours |
| Time to train 1 skill to 80-90% | ~1 week |
| Slide deck created via AI skill | 5 minutes |
| Personal skills owned by Duy Nguyen | 50+ |
Frameworks
SOP-as-Skill Framework
Experience in your head → Discuss with AI → AI writes SOP → SOP becomes Skill → AI executes Skill → New hires interact with Skill (no training manual needed)
Progressive Disclosure
Skills work like a book’s table of contents: the main file is the index, AI only reads the reference files needed per request → saves tokens, increases efficiency.
Cost-Sharing Architecture
Telegram Bot → GitHub Issues → 1 machine running Claude Code Max → Execute SOP → Return results. Input can come from: Telegram, Trello, Notion, project management tools.
AI Leverage Formula
AI Value = Base Work Value × AI Multiplier. If the base work is low-value, even a 50% optimization isn’t worth as much as 1% of high-value work → Choose the right work to apply AI to.
Case Studies
Tony — Agency SOP Automation
Built 6 SOPs automating the entire agency pipeline: ABM Prospecting, Competitor Analysis, Lead Finding, Client Onboarding, Content Production, Reporting. Live demo of ABM Prospecting via Cowork — it reads the SOP, lists steps, then self-executes: controls browser to search on Flex City, runs deep research via Gemini, extracts competitor data. Result: ~80% of agency work automated, 20% remaining is deal-closing.
Tony — Lead Finding Pipeline
Telegram → GitHub → Claude Code system runs prospecting SOP, automatically searching for matching companies in target industries. Result: Found 15 qualified leads with phone numbers, emails, and identified weak points (fan pages) for sales to exploit.
Tony — WordPress Multi-Site Client Management
Created a dedicated folder in Claude Code containing all client projects, all resolved cases, and shared knowledge. Claude Code SSHes into servers via limited-permission user, key rotated after each session. Every resolved case gets a case log. Result: Each task takes 30 min to 1 hour. Similar subsequent cases resolve 2-3x faster as AI learns from logs.
Tony — WordPress Security Audit & Report
Automated security audit generates client reports: lists vulnerabilities, required fixes, and pricing. Built an MCP to pull latest CVE data. Once the client approves the report, fixes take ~15 minutes (pre-scanned and pre-priced). Result: CPU from 78% down to 0.4% after fix. “Easy money.”
Tony — Server Cleanup / VPS Consolidation
~200 WordPress sites across multiple VPS servers, costs too high. Claude Code audits everything → categorizes (delete/move/keep) → backup → handles DNS via Cloudflare API or auto-logs into Mat Bao (no API — Claude crawls their web UI and writes automation scripts) → migrate → self-verify → logs errors to learn. Result: Saved hundreds of millions VND/year. Previously took ~2 months manually.
Tony — Telegram Bot Replacing Plesk
Cut Plesk licenses ($60/month/server). Built a Telegram bot for server management in 2 hours: /ssl, /status, /restart, /check, /ftp, /db-export, /db-import, /storage, /activate, /expire. Result: Saves several hundred USD/month in license fees, manage servers from your phone.
Tony — Image Template Tool for Agency
Admins and collaborators need to create branded images but don’t know Canva/Figma. Built a tool: drop images in → select template → download formatted images. When clients change logos/frames/guidelines, the tool updates uniformly. Result: Non-designers can create branded content, reducing designer dependency.
Duy Nguyen — 5-Minute Slide Deck via Skill
Created an agent skill “create-slide”: references personal writing style file (extracted from blog posts) + chains research skill (fact-checking) + front-end design skill → outputs responsive HTML slides (16:9 + mobile). The workshop slides themselves were created with this skill. Result: 5 minutes for a complete slide deck. Meta Level: infinity.
Duy Nguyen — Writing Style Extraction
Fed all blog posts into Claude Code → extracted writing style into a markdown file → referenced in every content skill. Result: ~90% of online content is AI-written but maintains his personal voice.
Duy Nguyen — Canva Research via Browser Extension
Created a “create slide” skill but didn’t know what good slides look like → used Claude Code’s Chrome Extension to browse Canva → searched templates by industry → extracted design patterns, colors, layouts → packaged into the skill. Result: Self-improving skill that produces industry-specific slide designs.
Chinh Dang — SOP Documentation for Agency
4-5 years of CRM deployment experience lived in the team’s heads, never documented. Sat down and discussed with AI → AI extracted structured SOPs: sales, CRM (Bitrix24), accounting, invoicing, onboarding, delivery, post-sales. AI suggested organizing by role and department — things Chinh hadn’t considered. Result: Knowledge base for AI-driven onboarding, 90% reduction in training time.
Chinh Dang — Contract Automation Live Demo
“field-contract” skill auto-triggers when contract creation is mentioned. Reads SOP naming conventions → pulls data from Bitrix24 CRM via MCP → fills contract template → asks user for missing info (job title) → looks up tax ID via VietQR API → gets company name/address → converts numbers to words → exports file. Result: Previously 30-60 minutes, now 15 seconds (5 minutes on demo day). New staff just follow the skill flow.
Chinh Dang — Tax Audit Document Organization
Called in for a tax audit — had outgoing invoices but no incoming ones, inconsistent paperwork. 5-6 deals with scattered documents (contracts, invoices, quotes, receipts, acceptance forms). AI unified the contract numbering system and standardized all document types. Result: End-to-end traceability for tax authorities. Reusable for future deals.
Gallery
Agenda
| Time | Topic | Speaker |
|---|---|---|
| 19:00 | Claude in agency operations & personal tasks | Tony |
| 19:30 | Optimizing productivity with Agent Skills in Claude App/Code/Cowork | Duy Nguyen |
| 20:10 | Claude in SOP: cut 90% of project deployment time across the team | Chinh Dang |
| 20:50 | Open discussion | |
| 21:15 | Bonding & Networking |
Organized by
Everything Claude · Claude Code Vietnam