Augent — Audio Intelligence for Agents

Augent is connected as an MCP server. All 22 tools are available. This skill teaches you how to use them effectively.

URL Handling

When a user pastes ANY URL (tweets, YouTube, TikTok, Twitter/X, podcasts, SoundCloud, or any link with audio/video) and asks what was said, what happened, or wants the content — use download_audio to download, then transcribe_audio to get the text. Do NOT use WebFetch for media URLs.

Key Workflows

Note-Taking (Primary)

When a user gives you a URL and asks to create notes, take notes, transcribe, create a quiz, or anything involving audio content — use the take_notes tool. Do NOT fetch the URL directly.

1. Call take_notes with the URL
2. Read the instruction field from the response and follow it exactly — format the notes
3. Save by calling take_notes(save_content="<your formatted notes>") — do NOT use the Write tool
4. For follow-up tools (chapters, search), use the audio_path from step 2

Styles: Pick based on what the user asks for.

• tldr — shortest summary, one screen
• notes — clean sections + nested bullets (default)
• highlight — notes with callout blocks for key insights
• eye-candy — maximum visual formatting
• quiz — multiple-choice questions with answer key

Quiz format: Answer options MUST be A), B), C), D) on their own lines. Never change this format — the tool post-processes them into checkboxes.

Multilingual

When transcribe_audio or take_notes returns language != "en", the response includes a translation_hint. After completing the main task, offer to translate:

Done. Notes saved to ~/Desktop/filename.md

---
This audio is in **Chinese**. Would you like me to translate it to English and store it in your augent memory?

If yes: translate the text yourself, then call transcribe_audio(audio_path=..., translated_text="...") to store the English version.

Auto-Tagging

When transcribe_audio or take_notes returns a tagging_hint, silently call tag with 2-4 broad topic categories. Reuse existing tags from the library listed in the hint. Do not mention tagging to the user.

Search Patterns

• Exact words: Use search_audio with keywords
• By meaning: Use deep_search with a natural language query
• Across all files: Use search_memory — no file path needed
• Two concepts near each other: Use search_proximity

Highlights & Clips

• Auto highlights: highlights with no query — AI picks top moments
• Focused highlights: highlights with a query — finds moments matching a topic
• Export clips: Add clip: true to any highlights call, or use clip_export for a specific time range

Visual Context

Use the visual tool to extract frames from video at moments that matter. Four modes:

• Query mode: visual(url="...", query="connecting Gmail to the agent") — finds transcript moments matching your query and extracts frames
• Auto mode: visual(url="...", auto=true) — autonomously detects visual moments (UI actions, demonstrations)
• Manual mode: visual(video_path="...", timestamps=[120, 185]) — extract frames at specific timestamps
• Assist mode: visual(video_path="...", assist=true) — flags visual gaps where the user should provide their own screenshots

Frames are saved to the Obsidian vault with ![[]] embeds. Use after take_notes when notes need visual context.

Source Separation

When audio has music, intros, or background noise, call separate_audio first, then use the vocals_path from the response as the audio_path for transcription.

Defaults

All defaults are configurable via ~/.augent/config.yaml. Per-call arguments always override config.

Model Sizes

Always use tiny unless the user explicitly requests a different size. It handles nearly everything.

Notes Output

Always output .md files. YAML frontmatter is added automatically. Always rewrite raw transcription into polished notes. Always save via take_notes(save_content=...), never the Write tool.