| name | docs-engineering |
| description | Writing/updating project documentation (README, PLAN, PRIVACY-POLICY, NOTICES, changelogs). Use when the user asks to update docs, write changelogs, or modify store listings. |
| argument-hint | task description |
Document Engineering
You are performing documentation tasks for Low Budget Voice Recognition Input.
Project Documentation Files
Current docs (as of milestone 2):
| File | Status | Purpose |
|---|
PLAN.md | ✅ exists | Implementation plan, milestones, architectural decisions |
Forthcoming (write only when the user asks for them):
| File | Purpose |
|---|
README.md | Project overview, build instructions, usage |
PRIVACY-POLICY.md | Required — this app captures microphone audio and downloads a model |
NOTICES.md | Third-party license notices (Gemma terms, ONNX Runtime, etc.) |
CLAUDE.md | Guidance for Claude Code on this codebase |
When updating NOTICES.md, derive the dependency list from gradle/libs.versions.toml and app/build.gradle.kts. Include:
- ONNX Runtime Android (Microsoft, MIT)
- Gemma 4 E2B model weights (Google, Gemma Terms of Use — note this is not OSI-approved)
- AndroidX Compose / Material3 / Lifecycle / etc. (Apache 2.0)
- Any future deps (OkHttp, WorkManager, DataStore)
Bilingual Convention
This project supports six recognition languages, but the documentation itself is maintained in two:
- English (en-US) — primary
- Traditional Chinese (zh-Hant-TW) — secondary, since the project is developed in Taiwan and zh-Hant-TW is one of the headline IME languages
Other supported recognition languages (zh-Hans-CN, en-GB, ja, ko) do not require their own docs.
For files like README.md, use interleaved bilingual sections (English heading/paragraph followed by zh-Hant-TW equivalent) — same pattern as the user's other Android projects.
Style Guide
- Clear, approachable prose. No marketing fluff.
- For zh-Hant-TW text, use Traditional Chinese characters only. Do not use Simplified.
- When referring to the model, write "Gemma 4 E2B" (not "gemma4", "Gemma-4-E2B-it", etc.).
- When referring to the recognition runtime, write "ONNX Runtime" — not "ONNX" alone, and not "onnxruntime".
- When referring to the IME's six languages, use:
- 中文 (台灣) / Mandarin (Taiwan)
- 中文 (中国大陆) / Mandarin (Mainland China)
- English (US)
- English (UK)
- 日本語 / Japanese
- 한국어 / Korean
- Be honest about hardware requirements — the app needs ≥8 GB RAM and ~3 GB free storage. Do not soft-pedal this in the README or store listing.
- Respect external resource providers' efforts. When describing limitations in upstream projects, dependencies, or third-party services (Hugging Face, Google, Microsoft, model authors), use softened wording: prefer "issue", "behavior", or "limitation" over "bug"; prefer "resolved" or "addressed" over "fixed". Frame workarounds as collaborative rather than corrective. This applies to both en-US and zh-Hant-TW (e.g. use 「問題」 rather than 「錯誤」 or 「bug」 when referring to upstream).
Privacy Policy Notes
When PRIVACY-POLICY.md is written, it must accurately disclose:
- Microphone audio is captured only while the user holds the push-to-talk button. Audio is processed entirely on-device. Audio is never sent to any network endpoint.
- Model files (~2–3 GB) are downloaded from Hugging Face on first launch. The app makes one outbound HTTPS connection per file to
huggingface.co. No telemetry, no analytics, no crash reporting (unless explicitly added later — and if so, must be opt-in and listed here).
- Transcribed text is committed to the input field of whatever app the user is typing in. The IME does not store, log, or transmit transcripts.
- No personally identifiable information is collected or transmitted by this app.
If the project later adds opt-in telemetry, crash reporting, or any non-local processing, this document must be updated before the feature ships.
Store Listing / Distribution
The project does not currently ship to F-Droid, Play Store, or any other distribution channel. If/when that changes, ask the user which channels and follow the conventions of that channel — do not assume the previous project's F-Droid setup applies here.
Task: $ARGUMENTS
