Whisper Extract

Whisper Extract

Transcribe audio or video with Whisper, then produce a .md file containing an AI summary followed by the complete literal transcript.

Prerequisite check

Before doing anything else, verify that Whisper is installed:

whisper --help > /dev/null 2>&1 && echo "OK" || echo "NOT FOUND"

If not found, tell the user to run:

pipx install openai-whisper
brew install ffmpeg   # if ffmpeg is missing

The turbo model requires openai-whisper >= 20240930; if the install is older, tell the user to run pipx upgrade openai-whisper.

Then stop and wait — do not proceed until Whisper is available.

Step 1: Gather required information

Ask the following in a single message if the user has not already provided them. Never ask more than once, and never ask for things already mentioned in the conversation.

Required:

• Audio/video file path — absolute or relative path to the file (mp3, mp4, wav, m4a, ogg, flac, webm, mkv, mov, mpeg, mpga, oga, wma).
• Language — spoken language in the recording. Examples: Spanish, English, Portuguese. If unsure, say "auto-detect" and Whisper will figure it out (slower).
• Recording context — a short description of what this is (e.g., "team meeting about Q3 roadmap", "interview with a candidate", "product demo call", "lecture on clean architecture"). This is used to write a better summary.

Optional (ask only if not obvious):

• Whisper model — default is turbo (recommended speed/accuracy balance). Options: tiny (fastest, less accurate), base, small, medium, turbo, large-v3 (most accurate, slower). Ask only if the user wants a different speed/accuracy trade-off.
• Output directory — where to save the .md file. Default: same directory as the audio file.
• Output language for summary — language for the summary and headings. Default: same as the recording language. If the user wants the summary in a different language, note it.

Wait for the user's answers before proceeding to Step 2.

Step 2: Transcribe with Whisper

Run Whisper on the provided file. Use the --output_format json flag to capture word-level timing and text cleanly, and --output_dir to control where the raw output goes. Clear /tmp/whisper-extract-temp first so leftovers from a previous run cannot be picked up.

whisper "<file_path>" --model <model> --language <language_code_or_auto> --output_format json --output_dir /tmp/whisper-extract-temp

Important: Always emit this as a single line — never split with \ continuations. A trailing space after \ is not a line continuation in zsh; it becomes an escaped space that Whisper receives as a second (empty) file path, causing ffmpeg to fail with Error opening input file .

Language codes: es for Spanish, en for English, pt for Portuguese, fr for French, etc. For auto-detect, omit --language entirely.

If the file is large (> 1 hour): Whisper will take several minutes. Tell the user:

"Starting transcription — this may take a few minutes depending on file length and model."

After the command completes, read the JSON file in /tmp/whisper-extract-temp/ whose basename matches the input file, and extract the text field. This is the full raw transcript.

If Whisper fails (file not found, unsupported format, ffmpeg missing), report the exact error and suggest a fix before continuing.

Step 3: Generate the summary

Given the full transcript text and the recording context provided by the user, produce a structured summary. Write it in the output language chosen in Step 1.

The summary must cover:

1. What this recording is about — one or two sentences.
2. Key topics discussed — bulleted list of the main themes or agenda items covered.
3. Key decisions or conclusions — if any were reached (skip this section if the recording is a lecture or monologue with no decisions).
4. Action items — concrete next steps mentioned, with owner if stated (skip if none mentioned).
5. Notable quotes or moments — 1-3 verbatim fragments that best capture the essence of the conversation (optional but highly recommended for interviews and meetings).

Keep the summary concise: aim for 150-300 words. Do not pad it.

Step 4: Write the .md file

Construct and save the output Markdown file.

File naming

Use this pattern: YYYY-MM-DD-<slugified-context>.md

Examples:

• 2026-04-14-team-meeting-q3-roadmap.md
• 2026-04-14-candidate-interview-backend.md
• 2026-04-14-lecture-clean-architecture.md

If today's date is available in context, use it. Otherwise, use the file's modification date via stat or just omit the date prefix and use the slugified context alone.

File structure

Use this exact template:

---
title: "<recording context>"
date: YYYY-MM-DD
model: <whisper model used>
language: <detected or specified language>
source: "<original filename>"
duration: "<approximate duration if available>"
---

# <Descriptive title based on context>

## Summary

<The summary generated in Step 3>

---

## Full Transcript

<The complete literal transcript from Whisper, paragraph-formatted>

Transcript formatting rules:

• Do NOT split the transcript into fake speaker turns unless Whisper detected them.
• Preserve the raw text exactly as Whisper returned it — do not paraphrase or clean up grammar.
• Wrap long monolithic output into readable paragraphs by inserting a blank line roughly every 10-15 sentences. This makes the file easier to navigate without altering the content.

Save the file to the chosen output directory using the current runtime's file-write capability.

Step 5: Confirm and show the result

After saving, show the user:

1. Full path to the .md file.
2. The summary section (so they can read it immediately without opening the file).
3. A one-line note about transcript length: Transcript: ~N words.

Do not dump the entire transcript in the chat — it's in the file. If the user wants to search or quote from the transcript, they can open the file.

Error handling

Tone

Write the summary in the same language as the recording (or the output language if specified). Be concise and factual — the summary serves as a quick reference, not a narrative essay. The transcript is the source of truth; the summary is the lens.