01
Taste context
What good looks like, what bad looks like, and what style to avoid.
brief
taste
scope
tests
Nikolay's vibe coding notes
Context is the new code.
AI can build fast. Your job is to give it a world where the right thing is obvious.
Estonian Business School
Cursor + Codex + practical build loops
Open with the thesis. This is not a tool tutorial. The useful part is how to think while using the tools.
About me
About me
RoleProduct manager @ Bolt
FocusInternal products and tools
BackgroundMostly B2B & enterprise
Coding1 year :D
Side quests10+ hobby projects
02
Use this to establish your perspective: PM, internal tools, enterprise workflows, and practical AI-assisted building rather than classic engineering background.
The actual bottleneck
The tool is fast. Context decides if it is useful.
Most failed AI prototypes are not caused by weak code. They are caused by missing taste, missing constraints, or missing validation.
02
Product context
Who it serves, which workflow matters, and what should stay out of scope.
03
Validation context
How the agent can prove the build works before it asks you to trust it.
03
Make the core frame concrete: taste, product, validation. Everything else in the talk maps back to these three.
Role shift
The AI writes more of the implementation. You own direction, tradeoffs, and the feedback loop.
Before
Write code
Search docs
Fix syntax
Wire pieces together
Now
Set context
Choose scope
Inspect output
Validate behavior
04
This is where you can compare Cursor and Codex briefly without turning it into a feature checklist.
Taste context
Give the AI taste, not just instructions.
References and anti-references are faster than a thousand adjectives.
Attachment 1
What I like
Use this as the north star for layout, density, tone, and interaction quality.
Attachment 2
What to avoid
Call out patterns you dislike: too generic, too playful, too dense, too empty, too SaaS-looking.
05
The anti-reference is the underrated move. It tells the AI where not to drift.
Product context
Discuss the brief before the first line gets written.
Who?Who is this for?
JobWhat moment or workflow matters?
ValueWhat changes if this works?
ScopeWhat is the smallest useful version?
No-goWhat should not be built yet?
SuccessHow will we know it is done?
06
Say that plan mode is useful after this conversation. First align on the brief, then let the agent plan.
Scope context
Even if you can build anything, start with one useful workflow.
Not the whole platformdashboards, auth, payments, admin, settings
Not just a pretty shellscreens that look good but do not prove value
Deliver one valuable workflowa user can complete it, see the result, and decide if it matters
07
This will land well in a hackathon. Push people away from giant dashboards and toward one memorable workflow.
Validation context
Give the agent ways to know whether the result works.
Agent build loop
Build
>
Review with tools
>
Fix
browser
tests
database
logs
Human final check
Use the prototype yourself.
Judge whether the workflow delivers value.
Then decide the next iteration.
08
Examples: browser access for UI, DB read access for saved data, Stripe read access for payment state, tests for expected behavior.
Parallel context
Make sure every prompt you send is easy for you later to read and understand.
Without structure
Five tabs open.
Three agents running.
No memory of what each one is doing.
You test the wrong thing.
Goal:
Context:
Problem to solve:
Expected impact:
Success metric:
What to test:
What not to touch:
09
Mention context switching. The task brief helps future-you understand what each agent is supposed to be doing.
Hackathon challenge
Today, build one working loop.
Briefwhat matters
Scopeone workflow
Build + validateagent checks with tools
Improvehuman validates value
10
Close with action. They should not leave thinking "AI is cool"; they should leave knowing exactly how to start their build.