AI B-Roll Finder: Matching Cutaway Footage to What You're Actually Saying

Most b-roll gets picked by scrubbing through a stock folder and eyeballing what looks close enough. An AI b-roll finder called b-roll-finder takes a different approach: it reads the transcript first, works out what each sentence is actually about, and only then goes looking for footage. The result is cutaways that match the claim being made, not the nearest keyword.

The project is built by developer Louise de Sadeleer and lives on GitHub at b-roll-finder. It ships as an agent skill for Claude Code, meaning there is no app to open and no dashboard to log into. It is a methodology file plus a set of CLI tools that an AI coding agent reads and follows.

How it matches b-roll for talking-head video

The workflow is plan-first, not source-first. Hand it a transcript, a video file, or an audio file, and it reads the whole thing before touching a single clip. It states the thesis of the piece, then works moment by moment, writing one sentence per beat describing what that line is actually about. A line like "Zaria from Duolingo made their social pop off" gets read as being about the person and the account she built, not the Duolingo logo. If nothing accurately illustrates the point, the beat gets dropped rather than padded with a stand-in clip.

Each moment then gets routed to whichever source can actually answer it correctly. The routing logic is the core idea of the tool:

• Receipts (something happening now, a claim, a complaint) route to tweets, headlines, or reviews, sorted by recency
• Entity moments (a named person, product, or historical event) route to the official or authoritative channel for that entity
• Concept moments (an abstract idea with no literal footage) route to custom motion graphics, built with Remotion
• Cultural or meme beats route to the user's own curated library, never auto-picked

Candidates are then scored on recency, source authority, relevance, and format fit before anything reaches the user. The tool narrows a large search down to a small contact sheet and stops there. Selection is left to the editor.

What it does after you pick

Once clips are chosen, the skill can cut them full-bleed and silent, then place each one on the exact spoken word using word-level timestamps from Whisper transcription, landing slightly after the word rather than before it, since early placement reads as a mistake and late placement reads as intentional. After placement, it pulls a frame from the midpoint of every beat and every cut point, tiles them into a contact sheet, and checks its own render for wrong shots or stray captions before handing the edit back. A manifest tracks which beats were approved so nothing quietly disappears in a later revision.

A built-in taste profile

Most automated b-roll tools stumble on curation, not sourcing. This one ships with a working taste profile called TASTE.md, built from scene-cut analysis of real published videos rather than a generic questionnaire. It encodes actual pacing numbers (roughly 26 to 28 cuts per minute, median shot length under two seconds), a topic-tagged list of trusted source channels, and guardrails like "if a plan is more than 60 percent website screenshots, it's wrong." The profile can be forked and rebuilt around a different creator's own published videos and YouTube subscriptions, which is the intended path for anyone who wants the sourcing to reflect their own channel rather than the default.

Who this is genuinely for

This tool is built for reference-rich talking-head video: podcast and interview intros, tutorial and explainer openings, competitor or "drama" videos with receipts, and listicle-style storytelling. The common thread is that the speaker names concrete things, people, products, companies, or events, which gives the agent something verifiable to source against. The more proper nouns and provable claims in a script, the better this works.

It is explicitly not built for pure mood or vibe montages, or music-only video where b-roll is a taste call rather than a factual match. The README is direct about this: it hands the wheel back to the editor in those cases rather than guessing.

Setup and what it actually costs to run

There is no signup and no API key. Setup means cloning the repository, registering SKILL.md with an agent such as Claude Code, and installing the CLI tools it drives: yt-dlp and ffmpeg for search, download, and compositing, imagemagick for contact sheets, and either mlx-whisper (Apple Silicon) or openai-whisper for word-level transcription. Remotion is an optional install for the motion-graphics route. A Chromium install is also used for headless screenshots of public pages, including handling consent walls automatically.

That setup bar means this is a tool for someone already comfortable running an agent skill and a handful of CLI dependencies, not a drag-and-drop editor plugin. It is aimed at solo creators, podcast editors, and small teams doing their own post-production who are tired of manually hunting stock footage for every named person or product mentioned in an intro.

Current traction

As of today, the repository shows 86 stars and 12 forks on GitHub, up from the 70 stars logged when it first surfaced as a fast riser. It was created in June 2026 and is licensed under MIT, so the source and the taste profile are both fully inspectable and forkable.

Source: github.com/louisedesadeleer/b-roll-finder