Build a personal English dictation app for my Windows PC in this project folder. Explain each handoff in plain English and do the implementation. Work in stages; verify each stage before adding features. GOAL A quick everyday tool: record speech, transcribe locally, insert the words where I was typing, and keep a short recoverable history. STAGE 1: INSPECT AND PREPARE Inspect the Windows version, CPU, physical RAM, free disk space, Python installations and microphone devices. Record the findings. Use a compatible 64-bit Python and CPU inference. Create an isolated environment; leave unrelated projects alone. Use ConwayResearch/Underdog-Bark-3.0-Clean. Read its current model card, licence, files and official bark_loader.py. Download the model, configuration and processor files needed by the selected loader. Record the model revision, file hashes and working dependency versions. Check the required audio-processing libraries, including librosa when the feature extractor requires it. Test imports in the same environment used by the launcher. STAGE 2: MAKE TRANSCRIPTION WORK Create a small window with Record, Stop, transcript text and a Copy button. List available microphones; let me select one. Use the capture device's actual sample rate and the model loader's required conversion. Handle stereo correctly. Load the model once. Show Preparing until it is ready. Keep recording disabled during preparation and transcription. Record audio, stop cleanly, and save a valid WAV before starting inference. Keep that file if transcription fails. Finalise the WAV header so another player can open it. Display the real transcript, the audio duration and elapsed transcription time. Handle empty recordings, silence and errors without inventing output. Keep the loaded model available for the next take. Use a worker for slow loading and inference so the window stays responsive. Let me set a recording-length limit. Warn before it is reached and save the recording at the limit. FIRST HANDOFF Show the actual executable or launcher I should open. Help me make one short recording and compare the words. Record the observed result before continuing. STAGE 3: ADD GLOBAL DICTATION Use Right Ctrl + Right Shift as the initial start/stop chord, with a setting to change it if it conflicts. Require a full key release between activations. A held chord must not trigger repeatedly. Show a small, visible Listening / Transcribing / Ready indicator without taking focus away from the app where I am typing. Remember the target window when recording starts. Insert the transcript only if that target is still the intended active window. Otherwise copy it and show clear feedback. Keep the transcript in history either way. Verify Unicode input, punctuation and line breaks in Notepad and the application I actually use. Report what happens in a field where insertion is blocked. Clicking X should hide the main window to the system tray. The tray menu must offer Open and Quit. Quit releases the microphone and shortcut. Starting twice should leave one running instance. STAGE 4: MAKE RECORDINGS RECOVERABLE Keep the latest ten recordings and ten transcripts by default. Let me change retention, including keeping a recording until I delete it. Pair each recording with its transcript using a stable ID and timestamp. Store them in a visible local folder, outside temporary build directories. Show the resolved path and provide an Open folder button. Handle Windows folder redirection and names with spaces. Verify that the actual folder opens in Explorer. Save transcript history atomically. Test reloading it after restart. Preserve user recordings and transcripts during repairs and updates. Delete older items only under the configured retention rule. If a file is locked, keep the app running and report the failed deletion. STAGE 5: PACKAGE AND VERIFY Build a normal Windows application with an obvious Start-menu launcher. Include every runtime file needed by the packaged app. Keep the EXE with its support folder if the package uses one. Complete all model downloads during setup. Restart with networking disabled and create a new transcript using the packaged app. Run the acceptance tests in this pack. Mark each PASS, FAIL or NOT RUN. Keep exact errors and repair the failed stage before adding features. DELIVER The installed launch path, source folder, data folder, requirements lock file, model revision, attribution, simple README and test log. State the actual checks completed and the checks that still need me to speak or click. Keep all audio processing local; omit cloud transcription and telemetry.