🤖 AI & IT Training
Building a Framework So AI Doesn’t Make Mistakes — Harness Engineering for Non-Developers
Same AI — so why does one person rewrite every output while another gets it right the first time? The difference isn’t the model, it’s the frame. Bundle your team’s rules into guidelines, skills, and review checks, and leave the 4-hour session with a working harness for your own tasks
⏱ Training hours4 hours
🏢 FormatOn-site delivery · Customized design
👤 InstructorJohn Kim
✅ Key outcomeFewer “wrong again” moments — once a mistake is fixed and written into the guidelines and checks, it doesn’t happen again
"We know AI is smart, but it keeps making the same mistakes." That’s what we hear most often from organizations that have finished company-wide AI training. It gets a form of address wrong, uses phrasing the company doesn’t use, and repeats this week the error you corrected last week. The cause isn’t the model. It’s the missing 'frame' around the model. The AI industry calls the work of building that frame harness engineering — just as you fit a horse with a harness, you fit AI with your company’s rules, memory, tools, and checks, so the same mistakes can’t recur. This introductory course is for non-developers in planning, sales, marketing, HR, and training. You won’t write a single line of code. The course material comes from real mistakes the instructor made while running a one-person consulting firm with AI, and from how those mistakes were turned into rules and checks so they don’t happen again. Each participant picks one task from their own work and ① writes the work rules as a guideline document the AI reads, ② packages a recurring task as a skill, ③ has the AI check its own output, and ④ feeds in deliberately wrong material to try to break that check and see whether it really catches errors. By the end, every participant leaves with a harness for their own work: '1 guideline page, 1 skill, and 1 review checklist.'
💻 Start learning this topic with an online course →Who is this course for?
This course is for non-developer practitioners in planning, sales, marketing, HR, and training who already use AI at work but rewrite its output every time; AI champions and innovation leads driving AI adoption in their departments; and team leaders who need to set team-level AI operating principles. No coding knowledge is required. Prior experience using generative AI such as ChatGPT or Claude at work at least once will help you move faster in the hands-on sessions.
What will you learn?
- Explain why the same AI gives different people different results — 'model vs. harness' — and break a harness into five layers: guidelines, memory, tools, checks, and permissions
- Write an instruction document that AI reads and follows, covering your team’s work rules (forms of address, terminology, prohibited expressions, document formats, reply commitments)
- Bundle one weekly recurring task into a skill set with defined inputs, steps, and output format, so the same quality comes out no matter who runs it
- Build a review checklist that lets AI inspect its own output, and deliberately feed it incorrect material to confirm the review actually catches it
- Create a permissions table separating what to delegate to AI from the buttons people keep to the end (send, pay, delete, publish externally)
Curriculum
Module 1: Same AI, different results — what is a harness? (about 40 min)
- Demo: run the same task (a customer reply email) twice, once without a harness and once with one, then compare the results
- The five layers of a harness — instructions · memory · tools (connections) · checks · permissions
- The difference between writing good prompts and building a framework — explaining every time vs. writing it down once
- Real case: the repeated mistakes a one-person company made working with AI, and how they were turned into rules
- Choose today’s practice task — one job you repeat every week and often revise
Module 2: Guidelines — turning your team’s rules into documents AI can read (about 60 min)
- Make a list of where AI goes wrong — forms of address, terminology, banned phrases, figures, document formats, reply commitments
- What makes a good instruction: rule + reason + how to apply it, with checkable statements instead of vague wording
- How to write what not to do: pair each banned phrase with its replacement
- Exercise: write a one-page instruction sheet for your work → add it to a Claude project (or Cowork working folder) and rerun the same task
- Rules for memory — what to store for the AI and what not to, and when an outdated rule turns into misinformation
Module 3: Skills — packaging recurring work into a reusable set (about 60 min)
- What a skill is: a task recipe with a defined input → procedure → output format
- Case study: a document skill that applies the same formatting rules every time to proposals, quotes, and training materials
- Hands-on: write one of your recurring tasks as a skill — required inputs, steps, final output format, and 1 example
- Principles for connecting tools — separating read from write access when connecting email, calendar, and drive to AI
- Run your own task with a colleague’s skill — confirm the result holds up when a different person runs it
Module 4: Review mechanisms — AI checks itself, people try to break it (about 50 min)
- Designing the review checklist — 5–7 items AI checks on its own before delivering its output
- 'A check isn’t proven until you’ve tried to break it' — feed in deliberately wrong material to test whether the checklist really catches it
- Real case: why a wrong result went out even though it passed the check — when the value the check looks at differs from the value on screen
- Hands-on: add a review checklist to your skill and try to break it with 3 incorrect samples
- Build a permissions table — what AI may do vs. the actions a human always keeps (send, pay, delete, publish externally)
Module 5: The operating loop and wrap-up (about 30 min)
- A harness is never finished — the operating loop of mistake → rule → check
- Sharing it with the team: moving personal prompts into shared team instruction and skill files
- Draft a one-page harness blueprint for your team — a list of rules, skills, and checks to add by next month
- Review today’s deliverables: 1 guideline page · 1 skill · 1 review checklist · permissions table
- Q&A and personal guidance on applying what you learn
Tools used in this course
Expected results / After the course
- 'Wrong again' happens less often — once a mistake is fixed, it goes into the instructions and the review checks, so it doesn’t come back
- Quality holds even when people change — a top performer’s know-how lives on as guideline and skill files the whole team shares, not as one person’s prompts
- The burden of reviewing AI output shrinks — AI does the first-pass check itself, and people only confirm the final buttons defined in the permissions table
- A next step after company-wide AI training — take home team-level operating principles for moving from using AI to delegating work to it
About the instructor
John KimInstructor John Kim, author of “B2B Marketing Design” and CEO of Nanum ConsultingView profile →
Related courses
Recommended online courses
Start with an Online Course
Before you book on-site training, preview the essentials of this topic in an online course.
AI · Online Courses
Claude Cowork · Hands-On Basics
An introductory hands-on course for building work tools that run on plain conversational instructions.
KRW 59,000 (≈ US$44)89,000View Course →
AI · Online Courses
Practical AI for Corporate Instructors
Use AI in your daily work, from lesson prep to content creation.
KRW 59,000 (≈ US$44)89,000View Course →
We tailor this course to your company
From diagnosis to execution — Nanum Consulting
FAQ
This course is for non-developer practitioners in planning, sales, marketing, HR, and training who already use AI at work but rewrite its output every time; AI champions and innovation leads driving AI adoption in their departments; and team leaders who need to set team-level AI operating principles. No coding knowledge is required. Prior experience using generative AI such as ChatGPT or Claude at work at least once will help you move faster in the hands-on sessions.
No. This course is designed for non-developer teams and involves no code at all. Guidelines, skills, and review checklists are all written as plain-language documents. Developer-oriented harnesses (automated check scripts, hooks, etc.) are introduced only as case studies of the underlying principles.
You need a personal laptop and a paid Claude account (a company Team account or a personal Pro account). For the hands-on practice, please bring one task you repeat every week and 1–2 samples of its output (use fictional material if security is a concern). We will send a pre-training guide before the session.
The principles of a harness (instructions, skills, review, permissions) are independent of any tool. During the pre-consultation stage, we set up the environment so the same exercises can run on your company’s approved generative AI (ChatGPT projects, in-house AI, etc.) using its instruction and knowledge-file features.
While earlier courses cover how to 'use' AI tools and how to 'build' automation, this course covers the framework for 'delegating' work to AI. It focuses less on tool features and more on writing rules, designing and deliberately breaking checks, and deciding what stays under human authority, which makes it a good next step after a company-wide AI training or a champion program.
Building a Framework So AI Doesn’t Make Mistakes — Harness Engineering for Non-Developers
Same AI — so why does one person rewrite every output while another gets it right the first time? The difference isn’t the model, it’s the frame. Bundle your team’s rules into guidelines, skills, and review checks, and leave the 4-hour session with a working harness for your own tasks
Training objectives
- Explain why the same AI gives different people different results — 'model vs. harness' — and break a harness into five layers: guidelines, memory, tools, checks, and permissions
- Write an instruction document that AI reads and follows, covering your team’s work rules (forms of address, terminology, prohibited expressions, document formats, reply commitments)
- Bundle one weekly recurring task into a skill set with defined inputs, steps, and output format, so the same quality comes out no matter who runs it
- Build a review checklist that lets AI inspect its own output, and deliberately feed it incorrect material to confirm the review actually catches it
- Create a permissions table separating what to delegate to AI from the buttons people keep to the end (send, pay, delete, publish externally)
Expected results
- 'Wrong again' happens less often — once a mistake is fixed, it goes into the instructions and the review checks, so it doesn’t come back
- Quality holds even when people change — a top performer’s know-how lives on as guideline and skill files the whole team shares, not as one person’s prompts
- The burden of reviewing AI output shrinks — AI does the first-pass check itself, and people only confirm the final buttons defined in the permissions table
- A next step after company-wide AI training — take home team-level operating principles for moving from using AI to delegating work to it
Curriculum
Module 1Module 1: Same AI, different results — what is a harness? (about 40 min)
- Demo: run the same task (a customer reply email) twice, once without a harness and once with one, then compare the results
- The five layers of a harness — instructions · memory · tools (connections) · checks · permissions
- The difference between writing good prompts and building a framework — explaining every time vs. writing it down once
- Real case: the repeated mistakes a one-person company made working with AI, and how they were turned into rules
- Choose today’s practice task — one job you repeat every week and often revise
Module 2Module 2: Guidelines — turning your team’s rules into documents AI can read (about 60 min)
- Make a list of where AI goes wrong — forms of address, terminology, banned phrases, figures, document formats, reply commitments
- What makes a good instruction: rule + reason + how to apply it, with checkable statements instead of vague wording
- How to write what not to do: pair each banned phrase with its replacement
- Exercise: write a one-page instruction sheet for your work → add it to a Claude project (or Cowork working folder) and rerun the same task
- Rules for memory — what to store for the AI and what not to, and when an outdated rule turns into misinformation
Module 3Module 3: Skills — packaging recurring work into a reusable set (about 60 min)
- What a skill is: a task recipe with a defined input → procedure → output format
- Case study: a document skill that applies the same formatting rules every time to proposals, quotes, and training materials
- Hands-on: write one of your recurring tasks as a skill — required inputs, steps, final output format, and 1 example
- Principles for connecting tools — separating read from write access when connecting email, calendar, and drive to AI
- Run your own task with a colleague’s skill — confirm the result holds up when a different person runs it
Module 4Module 4: Review mechanisms — AI checks itself, people try to break it (about 50 min)
- Designing the review checklist — 5–7 items AI checks on its own before delivering its output
- 'A check isn’t proven until you’ve tried to break it' — feed in deliberately wrong material to test whether the checklist really catches it
- Real case: why a wrong result went out even though it passed the check — when the value the check looks at differs from the value on screen
- Hands-on: add a review checklist to your skill and try to break it with 3 incorrect samples
- Build a permissions table — what AI may do vs. the actions a human always keeps (send, pay, delete, publish externally)
Module 5Module 5: The operating loop and wrap-up (about 30 min)
- A harness is never finished — the operating loop of mistake → rule → check
- Sharing it with the team: moving personal prompts into shared team instruction and skill files
- Draft a one-page harness blueprint for your team — a list of rules, skills, and checks to add by next month
- Review today’s deliverables: 1 guideline page · 1 skill · 1 review checklist · permissions table
- Q&A and personal guidance on applying what you learn
About the instructor
John Kim, Lead Instructor
10 years in industry · 15 years teaching (3,000+ lectures)
Education
· MBA (advanced major: Marketing), Korea University Graduate School of Business (graduated)
· Sungkyunkwan University Consultant Program (completed)
· Sogang University Graduate School of Business, Short-Term MBA (completed)
· Seoul National University Lifelong Education Instructor Development Program (completed)
· Chung-Ang University, B.A. in English Language and Literature (graduated)
Experience
· Current: CEO, Nanum Consulting
· Formerly at Hyosung Corp., Daeduck Electronics and Shinhan Diamond: overseas assignments and strategic planning (10+ years)
· Current partner instructor, Expert Consulting
· Current: Partner instructor, PSI Consulting
· (Current) Partner instructor, Korea SMEs and Startups Agency and SME Training Institute
Credentials and publications
· Certified consultant
· Certified teacher
· Completed official IDEO Design Thinking training
· B2B Marketing Design (1 Apr 2025)
· The Chinese Diamond Tool Market: Present and Future (co-authored, Korea Tool Association)
· Battles That Changed History and Innovative Weapons (in progress)