Can Claude Code edit a video by itself?

01 — The scorecard

What each editor actually did

Three edits of one 18:31 recording. Arm A is an autonomous Claude Code pipeline — word-level WhisperX transcript, FFmpeg, its own waveform refinement stage — with the instruction locked before the run and zero human intervention after. Arm B is TimeBolt's automated pass with default settings and no timeline edits. The reference edit is the version a human made by hand and would publish. It is the answer key, not a competing arm.

Take selection breaks down as 6 selections identical to the reference, 8 defensible alternatives, 1 clear error — a retained line missing its required lead. A different clean take is a choice, not a mistake; that rule was fixed in writing before scoring began.

Claude made better decisions and worse cuts. TimeBolt made safer cuts and left more mess. Neither was finished.

02 — The near-catastrophe

It nearly deleted the best 33 seconds, and it was certain

During setup, the Claude pipeline examined two windows in the recording — the passage naming competing tools, and an unscripted rant about handing internal footage to cloud AI — and concluded they were digitally silent. Not quiet. Zero amplitude. It ruled the transcript's 73 words there fabricated, built a guard against hallucinated text, and deleted 33 seconds.

Every step was locally sound: it probed the file, identified which audio channels carried voice, checked amplitude, confirmed across two windows, printed definitive, and engineered a fix. The speech was real. It lived on the channel pair the pipeline had already ruled out.

What caught it: a human reviewing the setup output before the locked run. One generic requirement was added — verify silence on every channel before calling a span fabricated — and the official run kept the content, correctly dropping only the profanity and filler inside it. Without that one rule, the strongest material in the recording would have shipped deleted, with a rigorous-looking log explaining why.

Automation's dangerous failure is not missing a cut. It is confidently removing the right thing for a wrong reason you can't hear.

03 — One word, predicted in advance

We named the word it would chop before it ran

Transcripts carry timing errors. In this recording's transcript, 16 words had end-timestamps that had absorbed the silence after them — one stretched to 20 seconds. We listed all 16 in the locked answer key, before the run, and predicted that cuts trusting those numbers would land in the wrong place.

The word “IP” was on that list. In Claude's finished edit, the sentence ends: “...they're going to get my.” The transcript said the word ended at one moment. The waveform knew it ended at another. The pipeline trusted the transcript.

Nineteen more chopped words followed the same pattern — 20 total, counted by two listeners, blind to which file they were hearing. TimeBolt's automated pass, which corrects every cut boundary against the physical signal, had 8.

A transcript tells you which words to remove. Only the waveform knows where they actually end.

04 — The sentence no transcript heard

Three editors kept it. Only one meant to.

WhisperX transcribed 741 words from the raw recording and missed a full spoken sentence, along with a two-minute stretch it returned as empty. The sentence exists in the audio. It exists in the finished edits. It does not exist in the transcript any AI editor was working from.

• The human kept it on purpose — because a person edits from what they hear, not from what a transcript claims.
• TimeBolt kept it by architecture — silence detection reads the physical signal, so audio present is audio retained. It never needed to know the words existed.
• Claude kept it by accident — a neighbouring word's malformed timestamp happened to stretch across that audio, dragging the sentence into a segment kept for other reasons. Verified in its own cut list: the sentence begins at 527.94s; the retained window runs 527.21–534.72s.

A transcript-driven pipeline cannot knowingly keep words it was never shown. On this file it kept them anyway — and could just as easily have lost them.

05 — Even the ruler bent

The measuring instrument failed three times

We scored these edits by re-transcribing each finished render. That instrument — the same transcription layer the entire AI-video workflow treats as ground truth — failed in three distinct ways during scoring:

• It missed roughly two minutes of real speech in the raw recording.
• It hallucinated “Thank you.” over about 20 seconds of real, intelligible content in one finished render.
• It emitted “...” over clearly audible speech in another.

Our first scoring pass, built on that instrument, wrongly charged TimeBolt's arm with three destroyed sections and nearly a minute of dead air. A higher-accuracy transcription pass retracted all of it. Both retractions are in the corrections log.

If the transcript can't be trusted to cut a video, it can't be trusted to grade one either.

06 — What the test cost us

Our own automation left duplicates. We shipped the fix in a day.

TimeBolt's automated pass came out of this test with seven duplicate takes and eight audible chopped words. That's on the scorecard above, unmodified, and it stays there. Within 24 hours we shipped a retake-detection update — v7.3.6, delivered automatically to every user — and re-ran the pass on the same recording.

Running the update at both settings isolated the cause. The algorithm produced the improvement, 8 down to 2. The more aggressive silence threshold in the middle configuration made things worse — it added three word shaves that appeared in both the transcript and the listening pass. Cutting harder is not cutting better.

Kept honest: these results sit in an addendum and never replace the locked test numbers. Claude's arm was frozen at lock and got no second run. Swapping in a post-test build would make this an advertisement instead of a test.

07 — What we got wrong

Seven corrections, published in full

Two AI systems scored this test independently and cross-examined each other's findings. Disputed claims were settled by artifact evidence — cut lists, timestamps, and blind listening — not by whoever argued harder. Seven claims were retracted or narrowed along the way, including our own scorer's confident mis-attribution of how the missing sentence survived, and its prediction that our padding change drove the boundary improvement. It didn't; the algorithm did.

Both scorers' full working files, every superseded claim, and the reasoning that overturned each one are published in the evidence set below.

A test where the scorers were never wrong is a test nobody checked.

Method

• Pre-registration. The answer key — 19 decision points, 16 scoreable, a junk inventory, 16 flagged timestamp defects, and five written predictions — was locked before either automated arm ran.
• Steelman. Arm A used the community-standard stack (Claude Code, WhisperX, FFmpeg) and built its own waveform refinement: 10ms RMS frames, 200ms boundary search, per-channel gating, padding, splice fades, full re-encode. It was not handicapped, and it did not cut on raw timestamps alone.
• Sequencing. The reference edit was completed and locked before any automated output was viewed. Setup instructions were iterated freely; the moment the run was locked, no human touched the output.
• Scoring. Divergences were bucketed as recoverable slop, destructive errors, or boundary errors — reported separately, never combined into one score. Repair cost is the weight.
• Listening. Two listeners, one blind to which render was playing, one agreed defect category, one pass per file.
• Disclosed bias. The reference edit was made in TimeBolt, because that is how we edit. Scoring measured the work required to reach publishable — not similarity to our timeline. The reference also contains deliberate stylistic beats no automation was asked to reproduce; those were excluded from scoring in both directions.

Evidence set

Locked answer key · scored results · cross-scorer tracker with all seven corrections · both automated outputs, untouched · the reference edit · full word-level transcripts of all four files · Arm A's cut list, config, and pipeline · the locked prompt, verbatim.

TimeBolt Vault · UMCheck · Bring Your Own Model