How to Make a Detective Game Where Spoken Choices Change Puzzle Rules

Build a detective game where spoken choices change puzzle rules with AI. Connect the new rules to evidence and conclusions.

Seele Editorial TeamUpdated November 18, 2026

At Greywater Quay, a witness says, “I crossed before the bell.” The ferry log places the muddy crossing at 21:00, but the phrase has two plausible meanings: the first strike at 20:00 or the end of its echo at 22:00. In this detective prototype, the player's spoken interpretation changes the time rule used by the puzzle. It must change the evidence state, not just the next line of dialogue.

Make interpretation change the case

An investigator studies muddy footprints by a harbor gate beneath a bell tower.
The bell tower, harbor gate, witness and footprints set up one contained timing mystery.

Use one case, three timestamps and two explicit readings. The first strike is 20:00, the crossing is 21:00, and the echo ends at 22:00. A muddy footprint supports that a crossing occurred at 21:00. It does not prove who crossed or whether the witness is truthful.

Spoken readingPuzzle boundaryResult for “before the bell”
The bell means its first strike20:0021:00 is not before it; account conflicts
The bell includes its final echo22:0021:00 is before it; account fits

Keep the choice finite. The player can say “first strike” or “last echo,” see the recognized phrase and confirm it. If speech is unavailable or uncertain, offer the same two buttons. Do not let recognition guess a rule silently.

Bind speech to the evidence rule

Store the selected interpretation as a case variable, then recalculate the timeline constraint. Show which boundary is active and why the footprint fits or conflicts. A reset returns the puzzle to unanswered; it should not erase collected clues. This gives the player a chance to compare readings without losing progress.

A detective compares two bell interpretations against a witness statement and crossing clue.
The witness clarifies the phrase while the investigator compares two clock boundaries with physical evidence.

Separate three states: transcript accepted, interpretation selected and conclusion supported. A clear transcript does not make the conclusion true. If the footprint is missing, either interpretation should remain unverified. If the choice changes, update the conclusion immediately and keep the previous selection available in the case log.

Replay the same evidence twice

The local rules replay used the 21:00 footprint and both bell readings. In the first version, speech was stored as dialogue, while the puzzle kept the same fixed rule; both readings were accepted without changing the evidence result. The revised rule used 20:00 for “first strike” and 22:00 for “last echo.” The first reading conflicted with the witness's “before” claim; the second fit it. With the footprint removed, neither conclusion was supported.

Replay caseDialogue-only versionRule-bound version
First strike, crossing at 21:00No changeStatement conflicts
Last echo, crossing at 21:00No changeStatement fits
Missing footprintChoice still acceptedNo conclusion supported

This deterministic local model checks one fictional rule, not speech recognition accuracy or a playable detective game. Next, test whether players understand the active boundary before adding suspects or another case.

At dawn, an investigator connects footprints, ferry records and the bell timeline.
The case board links the footprint, ferry crossing and selected bell boundary in one final review.

Start with one spoken interpretation

Ask an AI game maker for a small evidence puzzle with a fixed transcript choice. Keep the microphone optional, show the recognized words and let the player confirm before the rule changes. Add suspects only after the same evidence produces a clear, explainable result under both readings.

Create a detective game called The Greywater Bell. A witness says “I crossed before the bell.” A footprint places the crossing at 21:00. Let the player choose by voice or buttons whether “the bell” means its first strike at 20:00 or its final echo at 22:00. Show and confirm the recognized choice. Recalculate whether 21:00 is before the selected boundary, update the case conclusion, and log the active rule. If the footprint is missing, do not support either conclusion. Add a reset that preserves collected evidence. Never treat the interpretation alone as proof of identity or guilt.

Start a detective-game prototype in SEELE

Frequently asked questions

Should every spoken choice change a puzzle rule?

No. Use this mechanic when the phrase has two evidence-backed readings and players can see how each changes a constraint. Routine character dialogue can remain flavor. Keep the rule choice explicit, bounded and reversible so an unclear transcript does not silently rewrite a case or turn a deduction into a speech-recognition test.

What if voice recognition hears the wrong interpretation?

Show the recognized words and ask for confirmation before applying the rule. Provide buttons for the same finite choices, and let players correct the transcript without losing clues. The puzzle should depend on the confirmed interpretation, not on microphone access, accent or an unreviewed recognition guess.

Does a matching timeline prove that a suspect is guilty?

No. It only shows whether one statement fits the chosen time boundary and the collected clue. Keep identity, motive and guilt as separate conclusions that need their own evidence. This first prototype tests interpretation and timeline logic, so do not imply that a consistent time alone solves the whole case.

SEELE AI

Turn this guide into a playable build.

Start with a focused brief, then iterate in the Seele AI Workspace.

Start creating