Carry a prompt all the way to something usable.
A practical record of connecting ideas, conversation, prototypes, and checks until an AI output returns to a real screen and a human decision.
RESEARCH NOTE01
Name what needs to be made first.
A prompt is not a replacement for a brief.
Before drawing a screen, I describe who is stuck, where they are stuck, and what outcome would actually help. A feature name alone leaves input, response, exceptions, and the finish line undefined.
I reduce the task to a scenario that can be reproduced in a browser, then ask AI for options. That makes the directions comparable and keeps a polished paragraph from masquerading as a finished product.
The prompt is not the artefact I want to preserve. The useful part is a problem statement another person can reread and accept against.
- 01
Context
Who is stuck and why
- 02
Scope
What to make now
- 03
Acceptance
What proves it is done
02
A prompt is a brief you can return to.
Keep context, constraints, and current decisions together.
Long conversations mix abandoned options with current decisions. I keep a short list of what is accepted, what is deliberately out of scope, and which evidence is still missing.
In this site, AGENTS.md, DESIGN.md, content, messages, and tests carry different responsibilities. Giving an agent that map is not dumping every file into context; it tells the agent where rules, copy, and evidence live.
The point of context is to reduce guessing, not to turn the article into a specification nobody can finish reading.
- 01
Problem
Where the friction is
- 02
Decision
Which path is accepted
- 03
Boundary
What stays out
- 04
Evidence
How to check
03
Build a narrow slice, then return to real data.
Plausible prose is not usable behaviour.
I start with the smallest complete slice: one entry point, one primary action, and one result I can revisit. I test the structure with little data, then put long copy, error states, and the real phone width back in.
That step often disproves a prompt that sounded reasonable. A distant button, a broken title, or the wrong order of information is the product answering back.
The first demo is therefore not the result. It is material for the next question.
Brief
Describe the problem and limits
Slice
Build one complete path
Screen
Use it and look again
04
Components are not the answer; the next action is.
I decide what to draw by looking at the reader’s next move.
AI is good at splitting a screen into cards, buttons, and sections. The harder work is rhythm: when to give direction, when to add detail, and when to let a person decide.
On KENLIN.CC, Work, Journal, and Photography are separate entrances with different reading speeds. In KIROKU, one sentence from Discord must return to a place where it can be organized, searched, and revisited.
A good structure does not keep announcing its components. It makes the next action feel obvious.
- Orient
- Show where the reader is
- Act
- Offer one clear next move
- Return
- Make the result reviewable
05
Turn guesses into checkable evidence.
A completion report is not acceptance.
I split acceptance into layers: Typecheck for types, tests for behaviour, the browser for real interaction, and screenshots for reading feel. Each answers a different question.
A trilingual site also needs the same responsibility and evidence to survive in Japanese, Traditional Chinese, and English—not merely a translated string.
The benefit is knowing what to repair instead of ending with a vague “it seems fine.”
| STEP | Typecheck | Test | Browser | Pixels |
|---|---|---|---|---|
| Data shape is valid | Included | Not included | Not included | Not included |
| Behaviour matches the contract | Not included | Included | Not included | Not included |
| Interaction works | Not included | Not included | Included | Not included |
| The page reads well | Not included | Not included | Not included | Included |
06
What remains is what I am willing to use.
AI can speed up implementation, but it cannot live with the result for me.
I do not need every prompt to succeed at once. I need to know which assumption failed, how to brief the next pass more precisely, and how the result returns to my work.
Across translation, coding, and this site, AI has been a thinking partner, an implementation hand, and another pair of review eyes. The role changes the acceptance method.
A product is not finished because AI says so. It lasts when I decide I can keep using it.
- Speed
- Accelerate repeatable work
- Judgement
- Keep direction and boundaries human
- Use
- Return to daily use