AI-Assisted Editing

What's actually implemented

AI-assisted editing in transcriptcut means specific, scoped actions you explicitly trigger — not an open-ended chat interface:

• Filler-word removal — flags candidates like "um", "uh", "you know", "basically" for one-click removal. Detection is algorithmic, not generative. An LLM call is only used to classify ambiguous single words ("like", "actually", "so", ...) in context, deciding whether a given instance is filler or semantically meaningful — it never proposes cuts on its own, and it never touches words that aren't already flagged as ambiguous candidates.
• Long-pause removal — detects long silences and creates cut operations, again algorithmically (no LLM involved).
• Speaker detection — labels who said what. The model only says who spoke when; the words and their timings still come from the transcript.
• Chapters and show notes — generated for the Publish tab, fully editable afterwards.
• Find highlights — picks standalone moments for Shorts.

All of these are explicit buttons. Filler words and pauses are flagged for you to review before anything is cut. Speakers, chapters, show notes and highlights never cut anything by themselves: you rename, edit, delete or render them yourself. See the Podcast Workflow guide for where each one lives.

Models return picks, not timestamps

For chapters and highlights, the model gets a numbered list of sentences and answers with sentence numbers and text, never with times. The app then takes the start and end from the sentences' own word timestamps and checks the rules (chapter spacing, clip length between 15 and 90 seconds, no overlaps) before saving anything. A chapter or clip can't start mid-word or point at a time that doesn't exist, however the model answers. Every response is schema-validated, and invalid output is discarded rather than applied.

What was removed, and why

An earlier version of this project had a free-text "Ask AI" command bar — you could type something like "remove the first 30 seconds" or "cut the pricing section", and a LangGraph pipeline (Director → Editor → Validator agents) would turn that into structured edit operations. It was built and verified end to end, then deliberately removed: an open text field that could propose arbitrary edits to a project was decided against.

The agent architecture and reasoning are kept in claude.md (sections 5–8) as a record of that design, and could become relevant again behind a more constrained interface — but nothing in the current app implements a free-text AI editing command.

The core agent principle, if this returns

Whether or not a natural-language interface comes back, the project's rule for AI + media is fixed:

Agents reason. Tools execute.

User → LLM → Edit Operations JSON → Validation → Application logic → FFmpeg

An LLM is never allowed to produce or execute raw FFmpeg/shell commands. It only ever produces structured, typed edit operations, which are validated (timestamps in range, start < end, referenced assets exist, no impossible combinations) before any FFmpeg process runs.

Model choice

Task	Model	Why
Transcription	whisper-1	Word-level timestamps, which transcript editing depends on
Speaker detection	gpt-4o-transcribe-diarize	Speaker-labeled segments; Whisper still provides the words
Filler words, chapters, show notes	gpt-4o-mini	Narrow classification and summarizing: a small, low-cost model is enough
Find highlights	gpt-5.6-luna	Judging what works as a standalone clip across a whole episode needs a stronger reasoning model

Each sits behind its own service module on the server, so a provider can be swapped without touching the rest of the app.