Before we start
Let our questions drive us.
Six questions. Hold them; we will answer every one.
Question 1
Ever spent weeks or months building something, then learned it was the wrong thing?
Question 2
Agile. Design Sprint. Lean. Double Diamond. Why so many names for the same thing?
Question 3
What if idea to working prototype took an afternoon, not a quarter?
Question 4
What if vibe-coding had a method, learnable in an afternoon, not a hundred videos?
Question 5
What if the same loop built a skill, a web app, a plugin, a CLI, a library, a Cloudflare Worker, even your own marketing AI?
Question 6
What if you could start today, whatever experience you bring?
AI-Augmented Problem-Solving Loop
Realistic PDCA
Ron Michael Khu · Full-Stack AI & DevOps Engineer @ LegalMatch
Part 1 · The one idea
It starts with a cost.
The cost
Months spent building the wrong thing.
Idea → Build → Launch → Learn. Learning comes last, and costs the most.
The promise
Learn the loop, build anything.
Master the loop; the medium stops mattering.
P D C A
Deming, 1950s. Still works.
P
Plan
The problem, and what "done" looks like.
D
Do
Build the smallest real test.
C
Check
Measure against the goal.
A
Act
Standardize what worked, or loop.
What it really is
A problem-solving loop, not a ritual.
- Do not bet the company on an untested plan.
- Run cheap experiments. Learn fast. Refine.
Part 3 · The reveal
Every framework is PDCA in a hat.
They all map to PDCA
| Framework | Is just PDCA |
|---|
| Design Sprint | Map/Sketch/Decide = Plan · Prototype = Do · Test = Check |
| Agile / Scrum | plan · build · review · retro |
| Lean Startup | Build · Measure · Learn |
| Double Diamond | Discover/Define · Develop · Deliver |
| Design Thinking | Empathize/Define · Ideate/Prototype · Test |
Part 4 · The famous one
The GV Design Sprint.
What it is
5 days
Map → Sketch → Decide → Prototype → Test
Rules and roles
- Together, alone · getting started beats being right · do not rely on willpower
- Decider · Facilitator · cross-functional team (4 to 7)
Why it worked · the 5 Ps
Not the sticky notes.
- People: the team, and who you hire and keep.
- Process: the steps and the timeboxes.
- Purpose: the one problem you are solving.
- Platform: the tools and the outputs.
- Place / Pace: the environment and the tempo.
The 5 Ms (ops twin)
- Manpower: people and skills.
- Method: how the work gets done.
- Machine: the tools and equipment.
- Material: the inputs (data, feedback, briefs).
- Measurement: how you check results.
The uncomfortable truth
Process is portable. People and culture are not.
Copy the steps, you still cannot copy the team.
AI changes the math
- AI lowers the floor: you do not need Google's team to start.
- Experience still compounds: it raises the ceiling.
Part 5 · Before you vibe-code
What AI will not do for you.
The framing
AI makes building cheap.
It does not make judgment cheap.
Plan fundamentals
- Own the problem and what "done" means.
- Scope it: marriage vs fling.
- Skip this and AI builds the wrong thing faster.
Do fundamentals
- Prompting is spec-writing. Vague in, vague out.
- Keep a map of the parts: frontend, backend, data, auth, hosting, secrets.
- Do not leak: no API keys or customer PII in a prompt or a public repo.
Check fundamentals
AI is confidently wrong.
If you cannot verify it, you cannot ship it.
Act fundamentals
- Version control so you can roll back.
- Know your costs and limits.
Vibe-code the Do.
Co-pilot the Plan and the Check.
Part 6 · The AI age
Everything speeds up.
The classic sprint breaks
- Build-time → near zero.
- 40 hours of senior calendars = friction.
- Leans on team maturity.
Months → 5 days → 50 minutes.
Part 6½ · The trap
The better-mousetrap trap.
Emerson, 1871
"Build a better mousetrap, and the world will beat a path to your door."
the idea every builder wants to believe
It is not true
- 4,000+ mousetrap patents exist.
- The $1 snap trap still owns ~80%.
- People want the mouse gone, not a prettier trap.
Why "better" loses
- No distribution.
- Switching costs: needs 10x, not 20% (Thiel).
- Wrong problem.
And AI makes it worse: anyone can build a polished one in an hour.
The loop is the antidote
- Plan: right problem? who switches?
- Check: 10x? distribution? will they leave what they have?
Pick a real edge
- Cheaper (Temu, Backblaze): undercut on raw price
- Faster (Stripe, AWS): remove friction, instant infra
- Prettier (Notion, Linear): consumer-grade UX
"Better" is not a strategy. Pick an axis.
Part 7 · The method
Realistic PDCA.
Plan
AI brain-dump + scope
- Dump raw thoughts into an LLM; refine the goal and the real pain.
- Decide the relationship: marriage vs fling, before you build.
Do
AI artifacts in minutes
UI, script, draft, or agent. Minutes, not days.
Check · synthetic critics
- Skeptical CFO: pay or switch? 10x?
- Security auditor
- Impatient user
Then 2 to 3 real checks. Ego-free.
Act
Feed the critique back. Re-run.
Part 8 · The exercise
Pulse Prototyping.
Theory's over
Now we run the loop.
Grab paper, pencil, phone, or laptop. The clock starts now.
The five pulses
- 0 to 5: problem + 1 metric.
- 5 to 18: AI solutions + Impact-vs-Feasibility matrix.
- 18 to 32: AI wireframe / spec.
- 32 to 42: AI critique + silent dot-vote.
- 42 to 50: lock the winner + next actions.
Paper or digital, same loop
- Pulse 1 problem + metric: notebook, or any text app.
- Pulse 2 solutions + matrix: ChatGPT / Claude, or prompts by hand.
- Pulse 3 wireframe: fold paper into 3 panels and sketch, or generate UI.
- Pulse 4 critique + vote: pen checkmarks or sticker dots, or share screens.
- Pulse 5 lock winner: circle the top features.
Use what you already have open. Do not download anything.
No boss in the room? Good.
- The clock facilitates.
- The dot-vote decides.
- AI removes ego: critique the AI's options, not each other.
Part 9 · The payload
One loop. Everything we build.
Universality
The medium changes. The loop does not.
One loop builds all of these
- a Claude skill / agent for a product owner
- a web app
- a browser plugin
- a text-based tool (CLI)
- a backend library
- a Cloudflare Worker
- your own marketing AI
Worked example · ai-growth-tools
- Plan: PO needs X → "fling: a one-page intake helper."
- Do: generate the skill + a sample run.
- Check: run on a real ticket; critique.
- Act: tighten the prompt; ship or loop.
Part 10 · Close
Bringing it home.
Classic vs AI-era
| Dimension | Classic 5-day | AI-era (50 min) |
|---|
| Duration | 40 hours | under an hour |
| Facilitation | human | clock + templates |
| Research | 5 live interviews | synthetic + 2 to 3 real |
| Prototyping | a full day | minutes |
| Ego / friction | high | low |
Takeaways
- It is all one loop: PDCA.
- AI collapses the loop to an hour.
- Scope the relationship first.
- Critique the AI, not the people.
- Fix the broken P or M, not the loop.
The six questions, answered
- Wrong thing? → Plan and Check.
- So many names? → all one loop.
- Fast? → the 50-minute Pulse.
- A method? → Realistic PDCA.
- Builds everything? → the payload.
- Start today? → the door is open.
Parting advice
Ethical, human-centered engineering.
- Build human-centered teams; nurture culture and empathy.
- Technology is only as good as the people who build and use it.
The door is open.
Bring what you have, and AI multiplies it.
Thank you · Q&A
Realistic PDCA
Run cheap loops. Let AI do the heavy lifting. Keep the humans for judgment.
Ron Michael Khu · @ LegalMatch