How One Host Cut an Hour of Editing Down to 5 Minutes with AI

Background : "Chaojie" is a host at Shandong Radio‑TV who manages five new‑media shows, producing close to 200 episodes annually. Each 60‑minute program requires extracting 5‑10 valuable short clips, a task that traditionally takes about an hour of re‑watching, timestamping, cutting, adding backgrounds and titles.

Early AI Experiments

He first adopted a speech‑recognition tool to transcribe each episode into a TXT file (full text) and an SRT file (subtitle with timestamps). By reading the TXT he could locate interesting segments without scrubbing the video, and the SRT gave precise time positions.

Using these files he built a simple "subtitle timestamp extraction tool" where he manually chose start/end times and entered short titles. The tool generated a standardized timestamp file and a title file, which a separate Python script fed to FFmpeg to batch‑cut the clips. This eliminated the need to watch the whole video again, but the process was still heavily manual.

Splitting the Pipeline

Because a single program that performed all steps was fragile—any error in one stage broke the whole flow—he refactored the workflow into three independent programs:

Step 01 : Cut the original video according to the timestamps.

Step 02 : Place each horizontal clip onto a unified vertical background suitable for short‑video platforms.

Step 03 : Overlay the short title onto the final clip.

Each step is launched with a single click, making debugging straightforward. The three‑step pipeline reduced the processing time from 30‑60 minutes to 3‑5 minutes and proved far more stable than a “one‑click” solution.

Integrated Tool

After the three steps were reliable, he combined them into a GUI tool. Users select the original video, the timestamp file, the title file, a background image and an output folder. The tool then batch‑cuts, composites to vertical format, and writes titles, while offering options for font colour, tags and progress monitoring. Interns can operate it without programming knowledge.

AI‑Assisted Highlight Selection

To automate the most time‑consuming decision—choosing which segments are worth publishing—he created a "prompt generator". It feeds the TXT and SRT into an LLM with a template that defines criteria such as topic relevance, knowledge value, completeness and shareability. The model returns 10‑15 candidate clips, each with start/end timestamps and a short title, which feed directly into the short‑video tool.

Full End‑to‑End Workflow

The final pipeline is: transcription → AI‑guided highlight selection → timestamp matching → short‑video generation. Human reviewers only verify that clips are complete, titles are accurate and content is suitable for publishing.

Extending to Searchable Content

Beyond short videos, the full transcripts are used to build a searchable website for the host’s medical and legal programs. AI helps generate article drafts from the transcripts, designs the site architecture, and populates a database that lets users search by disease, doctor, hospital or legal issue, then jump to the original video. Video files remain hosted on existing platforms to keep storage costs low.

Key Lessons

Chaojie’s approach illustrates a step‑by‑step, human‑in‑the‑loop methodology: start with the most annoying manual task, automate it, verify stability, then layer additional automation. AI acts as a development partner that handles repetitive work, while domain expertise remains essential for content quality and decision‑making.