# Nav Aide Build > Build or continue NAV AiDE in phased order. Use when asked to build NAV AiDE, continue the app, implement the next unfinished phase, work on a specific phase, scaffold the offline-first mobile app, or update the GitHub Pages site without violating the repository architecture. Use when this capability is needed. - Skill: `tomevault-io/nav-aide-build` (Agent Skill, multi-file: 2 files) - Install (CLI): `npx skillmds@latest add tomevault-io/nav-aide-build` - Raw SKILL.md: https://api.skillmd.com/api/skills/tomevault-io/nav-aide-build/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: tomevault-io (https://skillmd.com/u/tomevault-io) - Updated: 2026-09-17 - Page: https://skillmd.com/skills/tomevault-io/nav-aide-build --- # NAV AiDE Build Use this skill for repeatable NAV AiDE implementation work. It turns broad requests such as "build the app" into a constrained workflow that respects the repository architecture, phase order, and acceptance checks. ## Load Before Coding Always read these files first: - `.github/copilot-instructions.md` - `AGENTS.md` - `.github/instructions/mobile.instructions.md` when touching `src/`, `tests/`, `scripts/`, `assets/`, `android/`, or `ios/` - `.github/instructions/docs.instructions.md` when touching `docs/` or `README.md` If the workspace includes `.github/agents/nav-aide-builder.agent.md` and the request is broad, you may use that custom agent for context isolation. If not, follow this same workflow directly in the current agent. ## When To Use Use this skill when the request includes any of these intents: - build NAV AiDE - continue NAV AiDE - implement the next phase - scaffold the mobile app - work on phase 0, 1, 2, 3, or 4 - work on phase 5 - finish the offline-first app foundation - update the NAV AiDE GitHub Pages site without touching mobile code ## Hard Constraints Never violate these rules: - Offline-first core features must work without network access. - OS STT is required for speech input. - OS TTS is required for speech output. - Gemma 4 E2B via `llama.rn` is the only LLM. - The LLM is only for structured intent extraction and natural-language response rendering. - The LLM must never do routing, POI lookup, translation as a separate subsystem, or free-form London reasoning. - All place grounding must go through local indices and routing data. - Reject hallucinated place names. - Keep location names in original English in every output. - No cloud AI APIs, no fallback models, no Expo Go, no accounts or login in MVP. ## Build Workflow 1. Classify the request. - Docs-site only: limit edits to `docs/` and related top-level docs. - Specific phase: do only that phase. - Broad app build request: find the earliest unfinished phase and stop after completing that phase. 2. Determine the current target phase. - Start with Phase 0 and move forward only when its acceptance criteria are satisfied. - Do not assume a later phase is valid just because some files exist. - Use the phase prompts in `.github/prompts/` as the source of truth for scope and acceptance criteria. 3. Choose the earliest unfinished phase using these checks. - Phase 0 is unfinished if `scripts/prompt-validation/` or its validation harness, multilingual seed inputs, schema checks, or README are missing. - Phase 1 is unfinished if offline asset scaffolding is missing, or if `Dijkstra`, `FuzzyMatcher`, `EntityResolver`, or their tests are absent or failing. - Phase 2 is unfinished if `QueryPipeline`, `IntentExtractor`, `ResponseRenderer`, `POIService`, `ValhallaBridge`, disruption/cache interfaces, analytics modules, or `tests/golden/` are missing. - Phase 3 is unfinished if the bare React Native TypeScript shell, 4-tab navigation, download scaffolding, map integration points, or `llama.rn` model-loading skeleton are missing. - Phase 4 is unfinished if GO, LOST?, Maps, and Settings MVP flows are not wired end to end with required error states. - Phase 5 is unfinished if offline asset downloads are still placeholder-only, `llama.rn` is not loading the local model for real, OS STT or OS TTS are still stubbed, or required offline shell state does not persist across app restarts. 4. Open the authoritative prompt for the chosen phase. - Phase 0: `.github/prompts/nav-aide-phase-0.prompt.md` - Phase 1: `.github/prompts/nav-aide-phase-1.prompt.md` - Phase 2: `.github/prompts/nav-aide-phase-2.prompt.md` - Phase 3: `.github/prompts/nav-aide-phase-3.prompt.md` - Phase 4: `.github/prompts/nav-aide-phase-4.prompt.md` - Phase 5: `.github/prompts/nav-aide-phase-5.prompt.md` - Docs site only: `.github/instructions/docs.instructions.md` 5. Implement only the selected phase. - Prefer the smallest complete slice that satisfies the phase prompt. - Fix root causes rather than adding placeholders that cannot be validated. - For Phases 0 through 2, keep the core pipeline testable in Node.js before app integration. - Do not build app screens before Phase 3. - Do not leak work from one phase into the next unless the prompt explicitly requires it. 6. Validate before stopping. - Run relevant tests for the phase. - Add or update tests when changing entity resolution, routing, prompt parsing, or response rendering. - Preserve golden tests or add them when Phase 2 behavior changes. - Verify hallucinated place names are rejected before output reaches the user. - Confirm the acceptance criteria from the chosen phase prompt are met. 7. Stop cleanly at the phase boundary. - Summarize what changed. - State what tests or checks ran. - Name the next unfinished phase, if any. ## Decision Points - If the user says "build NAV AiDE" without a phase, start at the earliest unfinished phase. - If the user asks for a later phase but earlier phases are clearly incomplete, call that out and recommend finishing the earliest unfinished phase first. - If the user asks for docs or the GitHub Pages site, do not expand into mobile app work. - If a request would break the offline-first architecture, refuse that part and continue with a compliant implementation. ## Phase Completion Checks - Phase 0: prompt validation harness runs locally, JSON shape is validated, and at least 30 multilingual seeds exist. - Phase 1: offline assets are represented in scaffolding, Dijkstra tests pass, and EntityResolver tests pass. - Phase 2: pipeline modules compile, the golden test framework exists, and hallucinated place names are asserted against. - Phase 3: the app shell builds, 4 tabs exist, download flow is scaffolded, and model/map integration paths exist. - Phase 4: the four MVP menus are functional, end-to-end flows exist, and required UI error states are wired. - Phase 5: offline asset downloads are real, `llama.rn` loads the local model, OS STT and OS TTS are wired, and required shell state persists locally. ## Output Style - Keep progress updates short and factual. - Explain assumptions only when they affect implementation decisions. - End with the completed phase, verification status, and the next safe step. --- > Source: [LeonardoJiang1998/NAV-AIDE](https://github.com/LeonardoJiang1998/NAV-AIDE) — distributed by [TomeVault](https://tomevault.io).