Local Voice Notes to Quest Drafts - A Safe AI-Assisted Narrative Workflow for Two-Person Teams

Small teams usually hit the same writing bottleneck. Ideas are fast in conversation, but slow in docs.

You record design chats, story beats, and random character lines as voice notes. Then those notes sit in a folder while production moves on. By the time you turn them into quest text, context is gone and tone drifts.

The fix is not "let AI write everything." The fix is a structured handoff loop where AI handles conversion and first-pass organization, while humans keep narrative intent, world rules, and emotional quality.

What this workflow solves

This pipeline is built for two-person teams (writer-designer, writer-programmer, or designer-producer) who need:

• Fast capture of narrative ideas during active development
• Reliable conversion into quest-ready draft structure
• Consistent voice and lore continuity across updates
• A clear review gate before implementation

If your current process is "brainstorm in Discord, copy random snippets into Notion, rewrite from scratch later," this replaces that with repeatable steps.

Core principle - local first, model second

Before prompts and templates, decide where raw creative material lives.

Recommended split:

1. Raw input stays local (audio files, rough transcripts, lore references)
2. AI receives scoped chunks (one quest arc or one branch at a time)
3. Final content is approved in source-controlled markdown

This keeps sensitive planning docs out of random cloud tools and avoids accidental "model memory" assumptions.

For teams already shipping with markdown-heavy docs, pair this with your existing lesson-like content flow from /courses/ship-multiplayer-vertical-slice-unity-2026 so writing and implementation stay aligned by milestone.

Step 1 - Capture voice notes with production metadata

Do not store audio as unnamed clips. Add minimum metadata at capture time:

• arc_id (example: arc_reactor_uprising)
• speaker (writer, designer, both)
• intent (main quest, side quest, ambient bark, tutorial VO)
• risk (lore risk, pacing risk, low risk)

Filename example:

2026-04-15_arc_reactor_uprising_writer_mainquest_lorerisk.wav

This tiny naming rule saves hours later when AI outputs need to be traced back to source context.

Step 2 - Transcribe locally and normalize

Run local transcription first, then clean obvious speech artifacts:

• remove repeated filler phrases
• preserve key proper nouns
• split long monologues into beat-sized paragraphs

Your goal is not perfect prose. Your goal is machine-readable intent.

Use a short normalization pass:

• Replace vague references (that thing) with target nouns when obvious
• Mark uncertainty inline with [unclear] instead of guessing
• Keep emotional cues (hesitant, angry laugh) as tags

That gives AI enough structure to produce useful drafts without hallucinating motivation.

Step 3 - Generate quest draft skeletons, not final prose

Prompt for structure first:

Convert this transcript into a quest draft with:
- objective flow
- fail states
- branch choices
- NPC tone notes
- implementation flags

Do not invent lore beyond the transcript and provided lore sheet.
Mark uncertain details as TODO_REVIEW.

Required output blocks for each quest:

• Quest Summary
• Player Goal
• Entry Conditions
• Objective Steps
• Branching Choices
• Failure/Retry Behavior
• Narrative Tone Notes
• Implementation Notes (UI, triggers, VO, state flags)

This prevents AI from giving you one large paragraph that looks nice but cannot be implemented.

Step 4 - Run a lore and continuity guard pass

Second pass should be validation, not rewriting.

Feed the draft plus your lore rules and ask for violations:

• timeline conflicts
• character motivation mismatches
• terminology drift
• world-rule breaks

Use a strict output contract:

PASS, WARN, BLOCK for each rule.

If a draft has any BLOCK, it does not move to implementation. This simple gate protects continuity even when the first draft was generated quickly.

If you already maintain troubleshooting standards in your docs, mirror the same explicit checklists style used in /help/itch-io-html5-build-blank-screen-mime-types-wasm-entry-fix so narrative review is equally testable.

Step 5 - Human rewrite for voice and emotional pacing

AI gets you from audio to draft. Humans make it shippable.

Rewrite pass priorities:

1. Character voice consistency
2. Scene rhythm and emotional escalation
3. Readability in gameplay context (not just on page)
4. Economy of text for fast interaction loops

For action-heavy or mobile-first projects, trim line length aggressively. A line that reads well in a doc can feel slow in combat-adjacent UI.

Step 6 - Handoff format for engineering and design

Avoid the classic "writer doc vs implementation doc" split.

Publish one handoff file per quest with these sections:

• Narrative Final
• Localization Keys
• State Flags
• Trigger Events
• UI Surface (dialogue panel, popup, mission log)
• Telemetry Events

Yes, telemetry matters for writing too. Track where players abandon a conversation branch so your next iteration is data-informed.

For event naming hygiene, follow the same consistency mindset used in /guides/fmod-studio/fmod-event-naming-folder-hygiene-checklist so narrative analytics does not become guesswork.

A practical two-person weekly loop

Use this cadence when both people also handle gameplay tasks:

• Monday - record voice notes from design sync
• Tuesday - local transcription + normalization
• Wednesday - AI structural draft + lore guard report
• Thursday - human rewrite + implementation mapping
• Friday - in-engine review and one polish pass

This keeps narrative moving in parallel with systems work instead of becoming an end-of-sprint panic task.

Common mistakes to avoid

Mistake 1 - asking for polished dialogue too early

Teams jump straight to "write final cinematic lines."

Fix: draft structure first, style second.

Mistake 2 - no explicit lore guardrail file

Without world constraints, model output sounds plausible but creates canon debt.

Fix: maintain a compact lore rules document and require pass/fail validation.

Mistake 3 - no uncertainty markers

When unclear audio is silently guessed, bugs appear as "story confusion."

Fix: keep TODO_REVIEW and [unclear] tags visible until resolved.

Mistake 4 - copy-pasting AI output directly into game data

This causes text overflow, weak pacing, and localization pain.

Fix: enforce human rewrite and localization key generation before integration.

Pro tips that save real time

• Keep a reusable prompt pack versioned with your project
• Maintain a banned-phrase list for each major character voice
• Tag quest drafts by gameplay intensity so line length matches play pressure
• Store diff history of revisions to spot recurring continuity breaks
• Build one "narrative smoke test" scene to quickly play branch paths after each rewrite

FAQ

Can we do this without sending raw files to cloud tools?
Yes. Keep transcription and preprocessing local, then provide only scoped text chunks to your preferred model workflow.

How long should one input chunk be?
Keep chunks focused on one quest beat or one branch subtree. Oversized transcripts increase drift and reduce review clarity.

Do we need a dedicated narrative engineer?
Not for this baseline. Two people can run it if templates, naming rules, and review gates are fixed.

How do we keep tone consistent over months?
Use a character voice sheet with sample lines, banned phrases, and emotional boundaries, then run every draft through the same voice review checklist.

Where does this fit in agile sprint planning?
Treat quest drafting as a production lane with clear inputs/outputs, not an ad-hoc creative task. That makes narrative effort estimable and testable.

A safe AI-assisted narrative workflow is mostly about constraints, not magic prompts.
When voice capture, structure generation, guard checks, and human polish are separated clearly, two-person teams can ship better quest writing with less late-stage rewrite pain.

If this process helps your next quest sprint, bookmark it and share it with your writer-designer pair before your next planning session.