Transcript Organizer — version 001

What the project does

Transcript Organizer runs an FFmpeg build containing the Whisper filter through a typed PowerShell interface. It checks the Windows runtime, executable capabilities, input media, model, output path, GPU policy, timeout, and result contract before it treats a transcription as successful.

Each run performs one internal JSON transcription and writes a reusable output folder containing raw and normalized timestamped evidence, readable text, SRT subtitles, media metadata, sampled frames, hashes, provenance, and validation.

Quick start

1. Download the distribution ZIP and extract it to a writable directory.
2. Install PowerShell 7 and an FFmpeg build containing the Whisper filter.
3. Place ggml-model-whisper-medium.en-q5_0.bin in the extracted project directory or its models subdirectory.
4. Copy a speech-bearing file such as video.webm into that directory.
5. Open PowerShell 7 in the directory and run the video file as the first argument.

pwsh -File .\Invoke-TranscriptOrganizer.ps1 '.\video.webm'

Preview the resolved backend, model, output folder, media duration, timeout, and execution plan without starting inference:

pwsh -File .\Invoke-TranscriptOrganizer.ps1 '.\video.webm' -DryRun

The launcher automatically examines project-local, WinGet, PATH, and Program Files candidates and selects a paired FFmpeg/FFprobe installation with the Whisper filter. For an authoritative custom installation, edit FFmpegPath and FFprobePath in the script’s user-editable settings block.

See the first-transcription manual for the module API, configuration object, model setup, VAD, CPU-only mode, batching, and troubleshooting.

Download a source video with yt-dlp

If you are authorized to download the source, yt-dlp can save it with a filesystem-safe name. Replace <VIDEO_URL> with the actual page URL.

yt-dlp.exe -o "%(title)s.%(ext)s" --restrict-filenames "<VIDEO_URL>"

The downloaded file appears in the current directory. Use its exact filename as InputPath. Download only material you have permission to access and process.

Generated outputs

To produce a complete corpus through the module API:

$manifest = '.\module\TranscriptOrganizer.psd1'
Import-Module $manifest -Force

$configuration = New-TranscriptOrganizerConfiguration -ProjectDirectory $PWD

Export-TranscriptOrganizerCorpus -InputPath '.\video.webm' -OutputFolder '.\video-transcript' -Configuration $configuration

Validated production result

The supplied 50-minute conference recording was used as a full field test on the documented Windows target machine. The run used the explicit FFmpeg 8.1.2 backend and one GPU attempt. No CPU fallback or retranscription was required.

• Corpus validation report
• Corpus manifest
• Latest test output
• Implementation and acceptance report

Reusable coding-agent skills

Three successful field workflows were distilled into dated, self-contained SKILL.md prompts. Each skill documents the complete extraction-folder structure and can start from an arbitrary validated video corpus without reading this talk or another experiment.

Generated report examples

These screenshots were captured from the actual local HTML outputs after the shared documentation style was applied. They are not mockups.

Experiment results

Every workflow begins independently from the shared extraction corpus. Experiments do not inherit prose, design decisions, evidence selections, or conclusions from one another.

1. Faithful illustrated talk companion

   Tested whether the corpus alone could support a complete technical article.

   Result: a detailed source-only companion covering the prepared talk and material Q&A with eight local frames. Reusable skill.

2. Primary-source researched field guide

   Tested whether external research could add depth without obscuring the talk.

   Result: a field guide using ten primary or authoritative sources with report, claim, and inference boundaries preserved. Reusable skill.

3. Interactive concept map

   Tested whether linear transcript evidence could become an inspectable systems model.

   Result: 20 editable timestamp-audited relations across limits, feedback, controls, expertise, and learning.

4. Saturation tabletop workshop

   Tested whether the talk’s mechanisms could support team practice.

   Result: a 95-minute scenario with five injects, decision gates, observer rubric, debrief, and worksheet.

5. Saturation response runbook

   Tested whether conceptual material could become an operational incident workflow.

   Result: a detect, diagnose, contain, and recover process with guardrails and service-specific decision gates.

6. 75-minute teaching kit

   Tested whether the source could support instruction beyond slogan recall.

   Result: objectives, timed activities, exercises, misconception checks, quiz, and a separate answer key.

7. Engineering leadership brief

   Tested whether mechanisms could be reframed as allocation decisions and evidence requests.

   Result: a compact decision brief and six prompts ready for a leadership review.

8. Skeptical claim audit

   Tested the talk as an argument to interrogate rather than a set of facts to repeat.

   Result: 13 claims classified by support, assumptions, challenges, and bounded verdicts in HTML and CSV.

9. Multi-surface publishing kit

   Tested whether one source could yield distinct editorial formats without quote drift.

   Result: newsletter copy, five short posts, two contextual image cards, and an editorial risk checklist.

10. Full transcript navigator

    Tested whether every normalized segment could remain directly inspectable without a server.

    Result: local search, eight chapters, recovery-provenance filtering, frames, and timestamp links for all 1,013 segments. Reusable skill.

Open the experiment collection for the compact catalog and clean-room protocol.