diff --git a/.gitignore b/.gitignore
index 287bedc..179b052 100644
--- a/.gitignore
+++ b/.gitignore
@@ -2,8 +2,8 @@
 __pycache__/
 *.py[cod]
 .pytest_cache/
+.superpowers/
 
 # Reference clips are local operational data; keep the directory in Git only.
 voices/*
 !voices/.gitkeep
-
diff --git a/docs/design/chatterbox-studio-concept.png b/docs/design/chatterbox-studio-concept.png
new file mode 100644
index 0000000..ec96203
Binary files /dev/null and b/docs/design/chatterbox-studio-concept.png differ
diff --git a/docs/superpowers/specs/2026-08-08-voice-studio-design.md b/docs/superpowers/specs/2026-08-08-voice-studio-design.md
new file mode 100644
index 0000000..6d54673
--- /dev/null
+++ b/docs/superpowers/specs/2026-08-08-voice-studio-design.md
@@ -0,0 +1,64 @@
+# Chatterbox Voice Studio design
+
+![Approved Chatterbox Studio concept](../../design/chatterbox-studio-concept.png)
+
+## Goal
+
+Add a local-network web interface at `chatterbox.dimensionlab.net` where any
+LAN visitor can upload, list, select, delete, and use local reference clips to
+generate Chatterbox speech.
+
+## Experience
+
+The root route (`GET /`) serves a single-page studio. It uses a calm white,
+two-column desktop layout with a responsive stacked mobile layout:
+
+- The left voice-library panel lists stored clips with filename, duration,
+  format, and selection state. It provides an upload drop zone/button and an
+  explicit delete action with an in-page confirmation.
+- The right generation panel contains the selected-voice summary, text field,
+  2,000-character counter, speed input, Generate button, and generated-audio
+  player with WAV download.
+- Empty states tell visitors to upload a voice first. Generation, upload, and
+  delete show clear inline progress, success, and error states.
+
+No authentication is required: this is intentionally available to everyone on
+the local network. The service remains bound to loopback; DNS and Caddy define
+its local-network exposure.
+
+## Voice storage and API
+
+Voice files continue to live exclusively under `VOICES_DIR`.
+
+- `GET /v1/voices` returns safe metadata for each managed file: relative name,
+  byte size, duration, sample rate, and channel count.
+- `POST /v1/voices` accepts a multipart `file` upload. The server sanitizes
+  the filename, rejects paths and collisions, accepts WAV only in this first
+  release, validates that the decoded audio is nonempty, and saves inside
+  `VOICES_DIR` only.
+- `DELETE /v1/voices/{name}` deletes one validated, regular file below
+  `VOICES_DIR`; it rejects traversal, absolute paths, symlink escapes, and
+  missing files.
+- `POST /v1/audio/speech` keeps its current OpenAI-compatible contract. The
+  studio submits the selected relative filename as `voice` and renders the
+  returned WAV locally in the browser.
+
+The existing model lock continues to serialize synthesis. Voice-management
+operations do not load the model. Deleting a clip does not interrupt an active
+synthesis request because speech validates and resolves its reference path
+before handing it to the model manager.
+
+## Error handling and safety
+
+All API failures are JSON with a clear 4xx/5xx detail. The UI renders these as
+inline messages. Deletion requires a confirmation click but no login. The
+upload limit is 20 MiB. Uploaded audio is never committed to Git; the existing
+`voices/*` ignore rule remains in place.
+
+## Verification
+
+Automated tests cover list metadata, WAV-only and 20 MiB upload validation,
+filename/path containment, duplicate rejection, deletion behavior, and the
+existing speech contract. Browser QA covers upload, selection, deletion
+confirmation, speech generation/playback/download, desktop layout, and the
+responsive stacked mobile layout.