PPS / 5W3H: One Sentence → a Full Task Spec via Automatic Prompt Expansion

STEP 1 — Expansion (in the same prompt):
  One sentence {task}
  → AI fills in the 8 5W3H dimensions
  → ~800-1200 words of structured brief

STEP 2 — Execution (right after the expansion):
  The 5W3H spec
  → The task is done according to all 8 parameters
  → Final result

Both steps — in a single request.

The 8 dimensions:

Task: Write a description of a new Google Workspace plan — an internal brief for the content team

Prompt:

My task: write a description of the new "Google Workspace Pro" plan 
for small business owners.

First, expand this task into the 5W3H structure — fill in all 8 dimensions:

**What (Task):** The specific task and success criteria
**Why (Goal):** Why it's needed, what the constraints are
**Who (Who):** Who is writing and for whom
**When (When):** Time context and deadlines
**Where (Where):** Where it will be published
**How-to-do (Method):** Structure and sequence
**How-much (How much):** Length, number of sections, format
**How-feel (Tone):** Style and delivery

Then write the plan description using this specification.

Result: The model first shows the filled-in 5W3H table — it pins down the audience (sole proprietors and small businesses of 2-10 people), the goal (conversion to purchase), the tone (businesslike but jargon-free), the structure (problem → solution → benefits → CTA), the length (300-400 words). Then it does the task strictly by that spec — no guessing.

LLMs fill in gaps — just not the way you want. When a prompt is incomplete, the model doesn't stop and ask — it generates text, plugging in "average" values for everything you didn't specify. Tone? Neutral. Length? Whatever comes out. Audience? General. Each model plugs in its own defaults — hence the spread in results.

But LLMs are great at structuring. Asking a model to "imagine the full brief for this task" plays to its strengths. It has seen thousands of examples of good briefs, specs and requirements docs. It knows what usually gets left out. In effect, you're using the model's knowledge of "good prompts" to improve your own prompt.

5W3H is a gap checklist. The eight dimensions are exactly the parameters users most often leave unspecified. Who (for whom?) and How-feel (in what tone?) are the most forgotten. When the structure is explicit, the model can't ignore them. So instead of an "average" prompt interpretation, you get precisely specified behavior.

Control knobs: - Refine the dimensions before execution — ask to see the expansion on its own, check it, fix it, then say "now execute" - Lock in what's critical — if tone or length matters, put it in the original request ("keep it casual", "no more than 500 words"): the model will preserve it during expansion - Pin a dimension — if you know what Who should be, spell it out: "the audience is team leads aged 30-40 with no technical background"

My task: {a one-sentence task}

Expand it into the 5W3H structure — fill in all 8 dimensions:

**What (Task):** The specific task and success criteria (KPIs)
**Why (Goal):** Why it's needed, goals and constraints
**Who (Who):** The author's role and target audience
**When (When):** Deadlines, seasonality, time context
**Where (Where):** Platform, channel, publication context
**How-to-do (Method):** Method and sequence of steps
**How-much (How much):** Length, number of elements, format requirements
**How-feel (Tone):** Style, tone, level of delivery

Then do the task strictly according to this specification.

What to fill in: - {a one-sentence task} — literally one sentence. "Write a landing page for an online Excel course." "Put together a negotiation plan with a supplier." "Explain the new revenue KPI to the team."

🚀 Quick start — paste into the chat:

Here is the 5W3H prompting template. Adapt it to my task: {your task}. 
Ask me questions so you can fill in the fields better.

[paste the template above]

The LLM will ask about the audience, tone and format — because those exact fields (Who, How-feel, How-much) are most often empty and have the biggest impact on the quality of the result. It will take the structure from the template and fill it in for your task.

⚠️ Raw JSON doesn't work: If instead of readable text with headings you give the model a bare JSON object with the same 8 fields, the result is worse than a simple prompt — especially noticeable in English. The structure needs to be readable text with labels, not code.

⚠️ Overkill for simple tasks: For something like "translate this word" or "write one sentence," 8 dimensions are excessive. The method works where there's real ambiguity: long texts, analysis, planning, multi-layered tasks.

⚠️ Composite scores are misleading: Simple prompts often get high "technical" scores (structure, completeness). But the study showed this is an artifact — with no explicit constraints, there's nothing to violate. On actual alignment with user intent, structured prompts win.

⚠️ AI expansion is the lower bound: The experiment used expansion without any user review. If you look over and tweak the generated spec before execution, the result will be even better.

The team took 60 tasks across three domains — travel, business, technical explanations — and translated them into Chinese, English and Japanese. For each task they compared four prompting variants: a plain sentence, 5W3H as raw JSON, 5W3H as readable text, and AI expansion of the plain sentence into 5W3H. Three models (DeepSeek, Qwen, Kimi) generated the answers, and a separate DeepSeek instance scored the results on five criteria plus a special "intent alignment" metric.

In total: 2,160 responses. They compared not only quality but also the spread between models: how much a structured prompt makes the result predictable regardless of which model you use.

The most interesting result: a statistically zero difference between a manual 5W3H spec (written by an expert) and AI expansion of a single sentence. In none of the three languages did the difference reach significance. This matters: it used to seem that the main thing in structured prompting was the user's effort in filling in the details. It turns out AI handles that just as well.

A curious detail: low spread between models on simple prompts isn't a virtue — it's a trap. The researchers showed that the "consistency" of simple prompts comes from all three models hallucinating the same way in the absence of constraints — and getting the same high scores from an automated judge that simply has nothing to mark down.

The four experimental conditions (original):

Condition A — Simple Prompt (Baseline):
"Please write a 5-day Tokyo travel guide including major attractions, 
budget information, and transportation tips."
Temperature = 0.7

Condition C — PPS Natural Language Rendering:
Task Goal (What): Write a 5-day self-guided Tokyo travel guide; 
  success criteria: cover ≥5 attractions per day, include transport links
Purpose (Why): Help first-time Japan visitors plan independently; 
  exclude venues closed in spring 2024
Persona and Audience (Who): Travel blogger writing for urban 
  professionals aged 25-35
Timing (When): Spring travel season (March–May)
Platform/Context (Where): Personal travel blog; Japan
Method (How-to-do): Overview → daily itinerary → top attractions → 
  transport → practical tips
Quantitative Requirements (How-much): 1,200–1,500 words; 
  include cost comparison table
Tone and Style (How-feel): Enthusiastic, beginner-friendly, 
  first-person voice
Temperature = 0, seed = 42

Condition D — AI-Expanded 5W3H:
[User inputs Condition A text → AI authoring interface expands 
remaining 7 dimensions → produces ~1,000-word natural language 
spec identical in format to Condition C]
Temperature = 0, seed = 42

Context: The researchers compared four approaches: a simple prompt (A), raw JSON (B), structured 5W3H text (C) and AI expansion of the simple prompt into 5W3H (D). Condition D is the practical discovery: the user writes one sentence, and the AI produces a spec identical in format to the hand-written C.

💡 Adaptation: Two-step mode with a review

Instead of executing the task right away, show the spec first and give yourself a chance to adjust it:

My task: {one sentence}

Step 1: Expand it into 5W3H — fill in all 8 dimensions.
Show only the spec; don't do the task yet.
Wait for my confirmation.

[After the model replies: "Fix Who: the audience isn't marketers, 
it's solo founders with no team. Now execute."]

This is a practical implementation of what the study calls the "lower bound": AI expansion without review is the minimal version. With review, it's better.

🔧 Technique: Lock in the critical dimensions

If you know what matters to you, say it explicitly in the first sentence. The model will preserve it during expansion:

My task: write a LinkedIn post about the launch of a new 
product. Audience — B2B; tone — no emojis and no exclamation 
points; length — 200 words max.

Expand it into 5W3H, keeping these three parameters as hard constraints.
Then write the post.

The AI will fill in the unspecified dimensions (How-to-do, Why, When) on its own. The critical ones are already locked in.

🔧 Technique: 5W3H as a template for recurring tasks

If you regularly do the same type of task (weekly reports, customer replies, newsletter posts), fill in 5W3H by hand once and save it as a "master template":

My standard template for [task type]:

**What:** A weekly digest of tech-industry news
**Why:** Keep subscribers in the loop without overload; 
  no politics, no hype
**Who:** Author — an analyst; audience — product managers
**When:** Friday, 6:00 PM ET
**Where:** Substack newsletter, 3,000 subscribers
**How-to-do:** 3 news items → a short comment on each → takeaway
**How-much:** 300-400 words, no long quotes
**How-feel:** Neutral, analytical, no value judgments

Here is this week's material: {paste the news}
Write the digest using the template.

Write the spec once — use it every week. The result stays consistent from request to request.

Main paper: Does Structured Intent Representation Generalize? A Cross-Language, Cross-Model Empirical Study of 5W3H Prompting — Peng Gang (Huizhou Lateni AI Technology Co. / Huizhou University, 2025)

Earlier work by the same author: arXiv:2603.18976 — the original PPS study in Chinese

Related techniques: CO-STAR, CRISPE, ICIO (structured prompt frameworks from the Schulhoff et al. survey)

Auto-expansion service: lateni.com — the interface used for Condition D