From Vague Prompt to Usable Draft: A Worked AI-Writing Production Example

A useful AI writing example should show the gap between “the model produced text” and “the team produced a usable asset.” Those are not the same event.

This worked example starts with a deliberately vague request:

We are planning a small fantasy exhibition. Write an article that explains why visitors should care about the exhibition’s worldbuilding.

The request is plausible, but it lacks almost everything needed for a dependable result. There is no audience, source packet, factual boundary, format, or definition of “care.” If we submit it as-is, a fluent model will probably fill the gaps itself.

The goal of the exercise is to turn that rough idea into a publishable creative article without inventing facts about a real exhibition.

Background: what we actually know

Assume the project is hypothetical. The approved notes contain only these facts:

the exhibition presents an original fantasy setting;
visitors can see concept art, character notes, maps, and prop designs;
the editorial theme is “how a fictional world becomes understandable through repeated rules and details”;
there are no verified attendance numbers, awards, celebrity partners, ticket prices, opening dates, or survey results.

That source boundary immediately tells us what the article must not claim.

Error 1: asking for persuasion before defining evidence

A weak prompt might say:

Write a compelling article proving this exhibition is a must-see experience.

The phrase “proving” creates pressure to manufacture reasons. “Must-see” encourages unsupported superlatives. The model has no evidence about visitor satisfaction, scale, rarity, or commercial performance.

The first correction is editorial, not stylistic:

Write an explanatory article for readers interested in fantasy art and worldbuilding. Use only the approved notes. Explain what visitors can learn from seeing concept art, maps, character notes, and prop designs together. Do not claim first-hand attendance, popularity, awards, ticket value, or visitor reactions.

Now the task has a factual ceiling.

Error 2: asking for an article before deciding the reader question

“What should readers care about?” is still broad.

A better reader question is:

How can an exhibition make fictional worldbuilding visible instead of leaving it as abstract lore?

That question gives the article an engine. Each exhibit type can answer part of it:

maps show spatial rules;
character notes show social and personal constraints;
props show how rules reach material culture;
concept art shows recurring visual decisions.

The article is no longer a promotional cloud. It has an explanatory job.

Correction 1: build a small source packet

Before generating prose, convert the notes into a structured packet:

Audience
Fantasy readers, illustrators, game/animation fans, and emerging creators.

Reader outcome
Understand four ways a physical exhibition can reveal the construction of a fictional world.

Approved facts
Only the four exhibit categories and the stated editorial theme.

Allowed invention
Hypothetical examples clearly introduced as examples, such as “If a city forbids open flame…”

Forbidden invention
Visitor quotes, attendance, dates, prices, awards, artist biographies, production budget, claims that something is “the first,” “largest,” or “best.”

Desired structure
Conclusion first → four exhibit lenses → one hypothetical mini-example → practical viewing checklist.

This packet takes minutes to prepare and prevents much longer cleanup later.

Correction 2: ask for an outline before prose

The outline request can be narrow:

Propose three structurally different outlines. Each must answer the reader question. Do not add facts. Avoid a generic “Introduction / Benefits / Conclusion” shape.

Possible outline A:

Start with a map detail.
Explain “rules become visible.”
Move from space → people → objects → visual consistency.
End with a visitor observation checklist.

Outline B:

Start with the problem of lore dumps.
Contrast text-only lore with evidence in designed objects.
Use four exhibit stations as a progression.
End with a creator takeaway.

Outline C:

Start with a hypothetical city rule.
Trace the rule across a map, character choice, costume/prop, and color language.
Explain how repetition creates world credibility.
End by asking readers to look for contradictions.

For this brief, outline C is strongest because it converts an abstract principle into an observable chain.

Draft pass: use a hypothetical example honestly

We can now give the model one invented example, labeled as invented:

Hypothetical example: imagine a city where open flame is prohibited after sunset because certain creatures are attracted to heat.

The article can trace that single rule:

maps might show covered night routes;
character notes might record professions affected by the ban;
props might include shielded lantern housings;
concept art might repeat low-temperature blue lighting in public spaces.

Notice what the example does. It demonstrates a reasoning pattern without pretending those details are actually in the exhibition.

This is one of the safest uses of invention: a clearly marked teaching example that is not confused with evidence.

Resulting sample passage

A usable paragraph might read:

Worldbuilding becomes easier to inspect when one rule leaves several kinds of evidence. Imagine, purely as an example, a city that prohibits open flame after sunset. A map might privilege covered routes; character notes might show which jobs become dangerous at night; props might replace torches with enclosed lamps; concept art might repeat cooler public lighting. None of those details needs a paragraph explaining “the fire law.” Their consistency lets the audience infer the rule. An exhibition can make this process visible by placing different stages of design beside one another.

That paragraph is modest. It does not claim a visitor reaction or invent an actual exhibit detail. Yet it still teaches something.

Review pass 1: unsupported-claim audit

Now highlight every phrase that sounds factual.

Potential problem phrases:

“visitors consistently notice…”
“the exhibition proves…”
“most fantasy productions…”
“the designers spent years…”
“this rare collection…”

None of those is authorized by the source packet.

Replace with language tied to what is known:

“placing maps beside props can make…”
“the stated theme invites readers to look for…”
“one useful way to examine worldbuilding is…”

The review should not make the prose timid. It should make the authority honest.

Review pass 2: structure and repetition

Suppose the previous four articles on the site all used:

What it is
Why it matters
Benefits
Common mistakes
Conclusion

Even a good new draft will feel automated if it repeats that skeleton.

Our worked example instead uses a traced rule. The structure follows evidence across media rather than repeating a generic explainer format.

This matters in scaled content production. Repetition is not only duplicate wording; repeated argument choreography also makes a site feel synthetic.

Review pass 3: make the conclusion actionable

Do not end with “Worldbuilding is important.”

End with a viewing tool:

When looking at a fantasy exhibition, ask:

Which rule appears in more than one medium?
Where do maps, costumes, props, and character behavior agree?
Where do they contradict each other?
Which detail changes because of economics, climate, politics, or ritual?
Can you infer a rule before reading its explanation?

Now the article gives the visitor something to do.

Chinese localization: do not mirror every sentence

The Chinese version should preserve the same source boundary and reasoning chain, but it does not need to copy English syntax.

For example, English might say “worldbuilding becomes inspectable.” A literal Chinese rendering would be awkward. A natural localization can say “世界观不再只是设定说明，而变成可以从物件和规则里看出来的东西.”

The factual boundary remains identical:

the same approved exhibit categories;
the same hypothetical status of the flame example;
no added visitor reactions or commercial claims.

Localization is a second editorial pass, not a mechanical export.

Before-and-after summary

Stage	Weak version	Usable version
Goal	“make people care”	answer a reader question
Evidence	unspecified	locked approved notes
Invention	uncontrolled	hypothetical examples labeled
Structure	generic article	trace one rule across media
Review	“sounds good”	claim audit + structure audit
Localization	sentence translation	purpose-preserving rewrite

Transferable rule

The best improvement came before generation: we changed an emotional request—“make this compelling”—into an editorial system.

The repeatable pattern is:

List what is actually known.
Define the reader’s question.
Decide what may and may not be invented.
Generate multiple structures before drafting.
Use hypothetical examples only when clearly labeled.
Audit claims separately from style.
Localize the article’s purpose, not each sentence.

That workflow does not guarantee brilliant prose. It does something more valuable first: it creates a draft whose strengths and weaknesses can be inspected, corrected, and trusted.