How this became a real interest, not a buzzword
My entry point into AI work wasn't chatbots — it was video editing. I was manually clipping long recordings down to highlight reels, and the repetitive part (finding the good moments, cutting, syncing) was exactly the kind of task a pipeline could do better than a person doing it for the tenth time.
That led to combining Whisper for transcription with an LLM for identifying interesting segments, then FFmpeg to actually cut the footage automatically. None of these pieces are exotic on their own — the value was in wiring them together into something that ran without me babysitting each step.
More importantly, it changed the question I ask before starting anything. Instead of "can AI do this?" — which almost always has some technically-yes answer — I started asking "what repetitive step is actually costing me time right now?" That's a much better filter for deciding what's actually worth automating.
1. Transcribe
The raw recording goes through Whisper to produce a timestamped transcript — this is the layer everything downstream depends on being accurate.
2. Extract highlights
An LLM reads the transcript and flags timestamp ranges likely to work as standalone clips — this is where prompt design actually matters.
3. Cut automatically
FFmpeg takes the flagged ranges and renders individual clips, so the manual work left is a quick review pass, not scrubbing the whole source.
Prompt engineering, practically
"Find the good parts" gives an LLM nothing to grade against. Giving it a concrete definition — a joke landing, a clear answer to a question, a moment where tone shifts — gives it something to actually check for. Try switching between the two tabs below to see why that difference matters.
Vague in, vague out — nothing here is directly usable by a script. Someone still has to watch the whole thing to find "the middle."
Structured, parseable, and each segment comes with its own justification — this feeds straight into FFmpeg without a human re-reading the whole transcript first.
Lessons learned
Automate the boring 80%, not the judgment call. Full automation without a review step tends to ship mistakes confidently. I keep a lightweight human check before anything goes out publicly.
Structured output saves more time than clever prompts. Asking for JSON with a fixed schema turned "parsing AI text output" from a fragile regex problem into a one-line JSON.parse.
The mindset shift mattered more than any single tool. Moving from "can AI do this" to "what repetitive step is actually costing me time" changed which projects were worth building at all.
Have a repetitive workflow worth automating?
Get In Touch