Find Words, Phrases, And Custom Transcript Moments
These two Library tools find transcript-identifiable moments in the active Premiere sequence and turn them into reviewable duration markers. Use the fast offline search when you know the wording. Use the semantic search when the meaning or the boundary of a moment matters.
These workflows run in Automation Agent for Adobe Premiere. See The Script Library for ready-made tools and Set Up Your AI for the Runtime Pack and Workflow AI setup paths.
Fastest Path: Run The Library Tools
Open Library > Examples > Transcripts & Captions > Transcript Search & Review.
Both tools capture the sequence that is active when you click OK, analyze
only transcript words that are actually used in that edit, preview the result,
and ask before creating markers.
| What you need | Use this tool |
|---|---|
| Find known words or phrases | Find Words And Phrases |
| Run quickly and completely offline | Find Words And Phrases |
| Allow bounded spelling or transcription variants | Find Words And Phrases |
| Describe a concept instead of exact wording | Find Custom Moments |
| Find the complete instruction after a production cue | Find Custom Moments |
| Detect real names, location clues, or context-dependent language | Find Custom Moments |
Both tools require the Python Runtime Pack. Find Custom Moments additionally requires a configured Workflow AI provider. Find Words And Phrases does not call an AI provider or network service.
Video: A Real First Test
This recorded first-use test compares literal matching with semantic search on a long interview, including how findings and their evidence are reviewed in Premiere.
The duplicate-result case shown in the video is historical test evidence, not the intended current behavior. The current workflow deduplicates fuzzy windows that refer to the same transcript occurrence while retaining genuinely separate spoken occurrences.
Find Words And Phrases
Choose Find Words And Phrases when you already know what the transcript should contain. Enter one word or phrase per line and optionally add fixed review time before and after every match.
Choose A Match Profile
- Exact only (recommended) ignores case, punctuation, Unicode presentation, and equivalent word separators, but does not guess synonyms or paraphrases.
- Exact and close transcript variants also accepts tightly bounded spelling variants such as a plausible one-character transcription error.
- Exact and broader transcript variants (review carefully) permits a wider bounded edit profile and can produce more false positives.
These named profiles are more predictable than a percentage slider. Exact matches take precedence. Very short words and numbers have additional safeguards, and the report identifies every approximate match. When several fuzzy word windows describe the same spoken occurrence, the tool keeps only the strongest candidate: exact before approximate, then higher similarity and the tightest word range. Separate spoken occurrences remain separate even when their added review handles overlap.
For example, a search for Jennifer does not report jennifer,
so jennifer, and jennifer we as three moments. Those windows share the same
transcript word, so only the exact jennifer result remains. A separately spoken
jenifer can still appear as one approximate result.
For a cue such as BEEP BEEP, first check how Premiere transcribes several real
examples. Search the consistent spelling exactly when possible; use a tolerant
profile only when the transcript contains a small number of predictable
variants.
The confirmation reports exact, approximate, overlapping, and coverage counts. A successful rerun replaces only markers previously owned by this tool, after the new markers have been created and verified. If a valid search returns no matches, the tool asks before clearing an older owned result set.
Find Custom Moments
Choose Find Custom Moments when literal wording is not enough. Describe the moments in plain language, for example:
- production cues plus the instructions that follow
- statements that may reveal the real production location
- mentions of real names or confidential project terms
- language that needs editorial or audience-suitability review
The tool sends stable transcript-unit IDs and their text to your configured Workflow AI provider. It validates every returned ID locally and asks before creating markers. A second bounded provider pass may refine the accepted start and end using supplied word IDs; the provider never supplies timestamps.
Choose How Much Context The Provider Sees
Whole transcript sends the complete usable sequence transcript in one request. Choose it for Codex CLI, Claude Code, or another provider that can reliably handle the input size, especially when distant context, callbacks, speaker relationships, or story-wide meaning affect the answer. If the input exceeds the hard request envelope, the run stops before calling the provider and recommends sections.
Overlapping sections sends deterministic target sections with preceding and following context. Each transcript unit belongs to exactly one target section, which prevents duplicate ownership while still allowing a finding to end in the following overlap. This is suitable for smaller or local models and locally observable requests such as a cue followed by an instruction. It is weaker when the decision depends on information far outside the visible section.
There is intentionally no automatic provider-based choice because the workflow cannot read a provider's context-window size. The setup dialog states exactly which search instruction, optional project context, and transcript text are sent to the provider. With a local provider they stay local; with a cloud-backed CLI they leave the machine under that provider's terms.
Before any Premiere write, the confirmation shows finding counts by confidence, provider-call and analysis-request counts, overlaps, boundary status, and transcript coverage. A verified rerun replaces only markers owned by this tool. A valid empty result can clear those markers only after a separate confirmation.
Understand The Shared Limits
Both workflows depend on Premiere's existing transcripts. They cannot hear a word omitted from the transcript, inspect a waveform, recognize silent actions, or analyze what is visible in the frame. Missing transcripts, unsupported timeline items, and incomplete speech coverage are reported instead of silently guessed.
Find Custom Moments adds language understanding, but its markers are still review suggestions rather than a compliance guarantee. Review every proposed range before treating the search as complete. Unrelated sequence markers are never part of either tool's replacement process.
For the narrowest Premiere write scope, open Execution Permissions, enable restricted Project write access, and allow Active Sequence Markers. The captured sequence remains the permitted marker target for that script run.