Keigo Higashino–Style Whydunit Rewriter (whydunit)
"Question-Pursuit" is the name of the technique, not a product name. Externally always say "mystery rewrite"; internally you may use "question-pursuit" to refer to the specific action.
What This Skill Actually Does
In one sentence: Add nothing, only rearrange information.
Detective fiction did not invent information asymmetry; it merely made "concealment" itself the theme. So the rewrite action is:
Find the thing already hidden in the story → hide it later → have someone unaware tell it.
Adding a corpse to a story is the crudest approach, guaranteed to produce garbage. If the first thing that comes to mind before starting is "who died," you've already gone the wrong way.
Core axiom: The protagonist is hollow; the supporting character bears the weight. If the person being questioned gives no inner life, the reader's emotions must be caught by someone else; the one who catches them is the true protagonist of this piece, and that person must have paid a price and have a share in the complicity. If this chain breaks, the whole piece collapses.
Key distinction: Whydunit is not whodunit. Higashino's mystery is almost never "who is the killer," but "why." In The Devotion of Suspect X, the reader knows the killer early on; the suspense lies entirely in motive and cost. The goal of the rewrite is to extract the unasked "why" from the original text, not to fabricate a "who."
Workflow
Execute sequentially; do not skip feasibility determination.
Step Zero: Ask the User Which Output Mode
Unless the user has already specified the desired output in the request, do not start directly. Use one question to converge on four options:
| Mode | Output | Use Case |
|---|---|---|
| A Rewrite Plan | Seven-part analysis (feasibility / hidden element / info reordering table / weight-bearer / outline / sample chapter / self-check) | Teaching, lesson planning, article writing, checking structural validity first |
| B Detective Novel | Full novel text, default 5,000-word novella | Ready-to-read final product |
| C Radio Drama | Scene-by-scene full-element audio script (dialogue + emotion + music + ambient sound), ready to synthesize | Broadcast-quality radio drama MP3 |
| D Audiobook | Pure voiceover rewrite + fixed timbre | Single-narrator audiobook MP3 |
When B is chosen, ask for length: 3,000 (five beats, one cross-check) / 5,000 (eight beats, three cross-checks, recommended) / 10,000+ (chapter outline, dual plotlines).
The difference between C and D is "whether to include ambient sound," not "whether a script is needed"—both can directly synthesize an audio file. C is a radio drama, D is a single narrator.
When C or D is chosen, do not ask the user to paste API keys into the chat, and do not call the API on the user's behalf. Honestly explain the reasons:
- Keys need not go through the conversation—the synthesis tool reads them from environment variables or a page password field.
- The runtime environment may not allow voice endpoints (sandboxes, corporate proxies commonly return 403/timeout). First synthesize a short test passage to confirm the path works, then produce the full script.
Correct approach: produce the script; the user synthesizes in tools/audio-studio/ (locally or on Hugging Face Space). API details are in references/seed-audio-api.md.
A/B/C/D all share the same rewrite plan as the master template. C is not a conversion of B into a script; it re-executes from the plan—different medium, different writing.
Step One: Feasibility Gate
Not every text can be rewritten. Honestly saying "I can't rework this" is far more valuable than forcing a rewrite.
Check three gates in sequence:
| Gate | What to Ask | Fail Condition |
|---|---|---|
| Information Asymmetry | Does the text contain something one character knows that another doesn't? | Pure lyricism, pure exposition, pure procedural text—no party holds exclusive information |
| Moral Weight | Does the concealed thing involve an ethical dilemma? | The secret is trivial, like "who stole the pencil" with no weight |
| Unasked Blank | Is there a "why" the author deliberately chose not to pursue? | All motives are already explained by the author; no holes to dig into |
All three gates pass → proceed to Step Two. Any gate fails → stop, output a determination report, explaining which gate failed, what narrative structure the original belongs to, and what a forced rewrite would look like. Do not force a conversion just to deliver something.
The moral weight gate is the most easily overlooked and most fatal: wrapping Higashino's structure around a trivial ethical dilemma results in farce, not suspense.
Step Two: Information Reordering Table
Before writing any body text, create this table. It is the foundation of the entire rewrite.
| Fact | Who Knows It | When Original Tells the Reader | When Rewrite Tells the Reader | Via Whose Mouth |
List at least 5 facts. All the rewriting effort is in the "When Rewrite Tells the Reader" column—same material, only the order and delivery channel change.
Step Three: Weight-Bearer Assignment (The Most Critical Step in the Entire Workflow)
Not choosing a narrator, but selecting the true protagonist of this piece.
The original protagonist will be hollowed out in the rewrite—their inner life is no longer presented. The reader's emotions consequently have nowhere to land, and someone must catch them. The person who catches them is the new protagonist.
Do three things in order:
-
Find in the original text someone who was casually mentioned but actually paid a cost because of the main event. A clerk, a neighbor, a bystander, a junior. That cost is a debt the original owes them; the rewrite collects it. Creating a new character is unacceptable—a sign the original wasn't read thoroughly.
-
Determine whether they are a camera. Ask: does this character feel worse at the end of the story than at the beginning? No → reselect. Someone who watches the whole thing unscathed cannot carry this piece.
-
Give them a share of complicity. They cannot merely observe; they must be implicated—laughed along back then, held testimony that could ruin someone, didn't reach out. The share can be minimal, but it must exist. This share is the sole source of the emotional impact later.
After assignment, check the character's information boundary: anything they cannot access must enter the text via retelling, rumor, or physical evidence—not via omniscience.
See the core axiom section in references/rules.md.
Step Four: Structural Outline
Read references/rules.md, then arrange chapter entries according to the structural layer rules. Each chapter must note: perspective / information advanced in this chapter / chapter-end hook.
Keep chapter count at 5–8. Fewer than 5 cannot sustain information delay; more than 8 is over-engineering at the outline stage.
In the 3,000-word and 5,000-word modes, the chapter outline will be replaced by beat structures in Step Five—but this step is still required; it provides material for the beat structure.
Step Five: Drafting
First determine the length mode. The three modes work completely differently, with different validation scopes (consistent with the conversion table in references/rules.md):
| Mode | Word Count | Structure | Investigation Density | Validation Scope |
|---|---|---|---|---|
| Short Story | 3,000 | Five beats | One cross-check, may contain one wrong turn | Group A + B7 |
| Novella | 5,000 | Eight beats | Three cross-checks, including one wrong turn (recommended) | Group A + B5/B6/B7 + Group M |
| Novel | 10,000+ | Chapter outline from Step Four | Three or more, dual plotlines | All clauses |
Novel mode: Follow the chapter outline from Step Four; all structural clauses apply.
Novella mode (5,000 words, default recommendation): Chapter outline is discarded; use the eight-beat structure. This is the smallest length that can simultaneously accommodate three cross-checks and three layers of reversal.
| Beat | Function | Word Count |
|---|---|---|
| 1 | Open with a present-tense scene, introduce an anomalous piece of evidence, do not explain | 500 |
| 2 | Flashback: the moment the weight-bearer paid a price | 700 |
| 3 | Background information, deliver as concisely as possible; subtly include temperature details (M2) | 400 |
| 4 | First cross-check: question a person. Provide evidence visible to the reader | 600 |
| 5 | Second cross-check: examine an object. Reach a wrong conclusion—the original's standard reading (S10) | 700 |
| 6 | Third cross-check: compare the two conflicting accounts, overturn beat 5 | 800 |
| 7 | Recover the temperature detail from beat 3; third-layer reversal: warm the cold answer | 700 |
| 8 | Revelation + weight-bearer's two-way choice (M3), do not close | 600 |
The wrong conclusion in beat 5 is not filler. Without an eliminated wrong path, there is no reasoning, only telling.
Short story mode (3,000 words): Five-beat structure.
| Beat | Function | Word Count |
|---|---|---|
| 1 | Open with a present-tense scene, introduce an anomalous piece of evidence, do not explain | 600 |
| 2 | Flashback: the moment the weight-bearer paid a price | 800 |
| 3 | Background information, deliver as concisely as possible | 500 |
| 4 | Key events, all relayed via secondhand accounts, conflicting with each other | 500 |
| 5 | Revelation + weight-bearer's choice, do not close | 600 |
3,000 words cannot accommodate dual plotlines and chapter-by-chapter puzzles. Must sacrifice S6, S7, S8 and Group M; preserve the core axiom and S3, S4, S11.
This trade-off is not a compromise; it is a load-bearing test: the clauses that survive at 3,000 words are the foundation of the rule system. Those that don't fit prove they are artifacts of length, not essential structure.
In short story mode, only judge Group A plus B7 (clue forewarning). B7 is non-negotiable—it is the only element that can still prove "this is reasoning, not telling" in 3,000 words.
Step Six: Validator Self-Check
Read references/validator.md, then score each clause per the validation scope table above, providing evidence for each. Unpassed items must be reworked from Step Two or Three; do not patch at the sample chapter level.
Clause numbering follows only validator.md. rules.md uses S/C/M numbering for explanation; validator.md uses A/B/M numbering for determination. The two numbering systems are not interchangeable (e.g., rules S11 clue forewarning = validator B7; rules C1/C2/C3 core axiom = validator A1/A2/A3; validator Group C is the C-a/C-b discrimination check, unrelated to rules Group C). Do not mix them in the self-check.
Output Format
Always output in this structure:
# Question-Pursuit Rewrite Plan: "Original Title"
## I. Feasibility Determination
(Three gates individually: pass/fail + reason)
## II. The Thing Hidden in the Original Text
(One paragraph: what was concealed, why it has weight)
## III. Information Reordering Table
(Table)
## IV. Weight-Bearer Assignment
(Who chosen / what price they paid / their complicity share / information boundary)
## V. Structural Outline
(Novel: chapter outline; novella: eight-beat table; short story: five-beat table)
## VI. Body Text
(Short story/novella mode: complete body text; novel mode: sample chapter opening, 800–1,500 words)
## VII. Validator Self-Check
(Each clause + evidence + rework direction for failed items)
Audio Mode
When the user requests an audiobook, audio, or radio drama, read references/audio-mode.md.
Audio mode is not the same text in a different format; it is a rewrite. Reading the text script aloud loses clue forewarning and flashback layering; you must add back references and verbal signposts.
Reverse Operation
The user may request the opposite: flatten an existing mystery text—tell it chronologically, enter the protagonist's inner life, explain everything clearly.
This operation is not a joke; it is the load-bearing test of the rule set. If deleting a rule does not make the text worse, that rule is not load-bearing and should be removed from rules.md. When such a request is received, still output a rewrite plan and explicitly answer at the end: which rules were proven load-bearing.
Boundaries
- Original copyright: The output must be an original rewrite. You may quote extremely short original excerpts to illustrate change locations (15 characters or fewer). Do not copy entire paragraphs from the original, and do not produce a retelling that substitutes for reading the original. For works still under copyright, prioritize structural analysis over full rewriting; user-owned texts and public domain works may be rewritten freely.
- Real events: When the material comes from a real person or incident known to the user, remind them to anonymize, and do not present speculation as fact in the plan.
- Style attribution: The output is "an imitation of Higashino's structure," not a work by Higashino. Do not write "Higashino would write it this way" in the plan; write "per this structure."
- Sentence-level is not the focus: Imitating short sentences and plain description is easy, but plainness itself is not a hallmark, and no one would recognize it even if learned. All effort goes into the structural layer.
Reference Files
references/rules.md— Three-layer rule set (sentence/structural/motif); required reading at Step Fourreferences/validator.md— Validation checklist; required reading at Step Six; clause numbering follows this filereferences/worked-example.md— Full walkthrough of "Kong Yiji"; read it when unsure what granularity is expectedreferences/audio-mode.md— Rewriting rules for mode C / D; required readingreferences/seed-audio-api.md— Doubao audio generation model 1.0 API, segmentation thresholds, timbre