Finding the Problem: Start From One Concrete Annoyance

What natural-language programming actually is, how to source a topic from your own week, and how to decide how much version one should do.

No tools in this lesson. Two things only: shrink "it would be nice if a tool existed" into one person, one scenario and one task; then decide how far version one has to go to be enough.

Four terms in plain language

Term	Plain meaning	How it is used here
Product manager	The person who decides who is served, what comes first, and what "finished" means	That is you. AI does not make these calls
Vibe coding	Describe what you want in everyday language and let AI design, write and revise the code	You specify, it builds, you try it and give feedback
MVP	A version with few features that still solves one whole problem and can be tried by a real person	One main flow from input to result
Requirement, design, acceptance	Three questions: for whom, how it is used, and what counts as correct	Three lists, six sentences, one page

Vibe coding is not "say one sentence, get a good product". The real loop is: specify → look at the result → name the problem → revise. So the whole of lesson one goes into specifying.

Step one: mine the past week

Do not start with a product name. Start with a task. Ask four questions about the last seven days:

Which task do you repeat with almost the same steps every time?
Which step keeps breaking, getting skipped, or needing a manual check?
Which existing tool is far heavier than the small part you actually use?
If this tool existed today, who would want it immediately?

Then compress the answer into one sentence:

I want to help [who] in [which situation] turn [how it is done today] into [what concrete result].

Version	Workable?
I want to build a powerful AI productivity tool	No. No user, no scenario, no way to tell if it is done
Every time I get several material lists I hunt for duplicates by hand; I want to paste them once and get a clean copyable list	Yes. User, scenario, input and output are all present

Step two: pick a direction close to your work

Situation	What the work is	What version one can do
Event material lists	Merge several overlapping text lists	Paste text, drop blank lines and exact duplicates
Exam revision	Turn organised notes into practice cards	Enter question and answer, click to reveal
Pre-publish copy check	Self-check before publishing	Flag missing items against your own checklist
Product specifications	Choose a combination from given options	Generate a spec summary from pre-filled options
Trip preparation	Pack by scenario	Choose a scenario, get a tickable packing list
Glossary lookup	Search a small glossary you provide	Enter a keyword, show matching entries
Group activities	Random grouping or ordering	Enter fictional codes, group by rule
Pre-rename file check	Preview a consistent naming scheme	Show before/after names, change nothing on disk

Or ignore the list and take something from your own experience.

Step three: four questions to test the topic

Can I name one specific user and one specific situation?
Can I prepare a few sample inputs without sensitive data?
Can I describe a clear output?
Can I judge myself whether the result is right and useful?

All four need a yes. If the answer is only "smarter", "prettier" or "fully automatic", the problem is still too big — cut it down again.

A cheap pre-check: try solving it once with an existing tool or a single AI conversation. If once is enough, you do not need a tool; if you would repeat it, build it.

Step four: size version one with one question

If this feature is removed, can the user still complete the key task?

For a list-cleaning tool:

Must have in version one	Later
Input: one item per line, pasted text, prompt when empty	Import from Excel, PDF or images
Clean: drop blank lines and exact duplicates, show the count	Merge items that mean the same thing
Take the result: select and copy the cleaned list	Login, collaboration, cloud history, sync

Version one solves "clean a text list" and nothing else. A plain interface is fine; wrong rules or an unusable result means it is not done.

The point of an MVP is early use, not polish. After building, observe two things: did it remove the original step? Which necessary step is still missing? That is all the input you need for version two.

The point to catch: smaller is better. One person, one scenario, one task, one checkable answer. If the topic is too big, the checklist in lesson two cannot be finished and nothing gets built in lesson three.

No tools in this lesson. Two things only: shrink "it would be nice if a tool existed" into one person, one scenario and one task; then decide how far version one has to go to be enough.

Four terms in plain language

Term	Plain meaning	How it is used here
Product manager	The person who decides who is served, what comes first, and what "finished" means	That is you. AI does not make these calls
Vibe coding	Describe what you want in everyday language and let AI design, write and revise the code	You specify, it builds, you try it and give feedback
MVP	A version with few features that still solves one whole problem and can be tried by a real person	One main flow from input to result
Requirement, design, acceptance	Three questions: for whom, how it is used, and what counts as correct	Three lists, six sentences, one page

Vibe coding is not "say one sentence, get a good product". The real loop is: specify → look at the result → name the problem → revise. So the whole of lesson one goes into specifying.

Step one: mine the past week

Do not start with a product name. Start with a task. Ask four questions about the last seven days:

Which task do you repeat with almost the same steps every time?
Which step keeps breaking, getting skipped, or needing a manual check?
Which existing tool is far heavier than the small part you actually use?
If this tool existed today, who would want it immediately?

Then compress the answer into one sentence:

I want to help [who] in [which situation] turn [how it is done today] into [what concrete result].

Version	Workable?
I want to build a powerful AI productivity tool	No. No user, no scenario, no way to tell if it is done
Every time I get several material lists I hunt for duplicates by hand; I want to paste them once and get a clean copyable list	Yes. User, scenario, input and output are all present

Step two: pick a direction close to your work

Situation	What the work is	What version one can do
Event material lists	Merge several overlapping text lists	Paste text, drop blank lines and exact duplicates
Exam revision	Turn organised notes into practice cards	Enter question and answer, click to reveal
Pre-publish copy check	Self-check before publishing	Flag missing items against your own checklist
Product specifications	Choose a combination from given options	Generate a spec summary from pre-filled options
Trip preparation	Pack by scenario	Choose a scenario, get a tickable packing list
Glossary lookup	Search a small glossary you provide	Enter a keyword, show matching entries
Group activities	Random grouping or ordering	Enter fictional codes, group by rule
Pre-rename file check	Preview a consistent naming scheme	Show before/after names, change nothing on disk

Or ignore the list and take something from your own experience.

Step three: four questions to test the topic

Can I name one specific user and one specific situation?
Can I prepare a few sample inputs without sensitive data?
Can I describe a clear output?
Can I judge myself whether the result is right and useful?

All four need a yes. If the answer is only "smarter", "prettier" or "fully automatic", the problem is still too big — cut it down again.

A cheap pre-check: try solving it once with an existing tool or a single AI conversation. If once is enough, you do not need a tool; if you would repeat it, build it.

Step four: size version one with one question

If this feature is removed, can the user still complete the key task?

For a list-cleaning tool:

Must have in version one	Later
Input: one item per line, pasted text, prompt when empty	Import from Excel, PDF or images
Clean: drop blank lines and exact duplicates, show the count	Merge items that mean the same thing
Take the result: select and copy the cleaned list	Login, collaboration, cloud history, sync

Version one solves "clean a text list" and nothing else. A plain interface is fine; wrong rules or an unusable result means it is not done.

Four terms in plain language

Step one: mine the past week

Step two: pick a direction close to your work

Step three: four questions to test the topic

Step four: size version one with one question

Table of Contents

Finding the Problem: Start From One Concrete Annoyance

Four terms in plain language

Step one: mine the past week

Step two: pick a direction close to your work

Step three: four questions to test the topic

Step four: size version one with one question

Table of Contents