The proposal stays visible
Proposed in/out points, landed edges, drift, and the rule that moved each boundary remain side by side. Manual edits update the audit trail instead of erasing it.
CLIPJAM REVIEWLOCAL-FIRSTPYTHON 3.11+
Find the thirty seconds. Land the cut.
ClipJam turns long-form YouTube video into reviewable social clips and a stitched compilation. Gemini proposes the moment. Deterministic code decides where the blade lands. A human approves before anything renders.
Private repository · source access available on request
SELECTED CUTWhy they called it the Empire of the GreeksPROPOSED
SOURCE POSITION16:29.00 → 17:00.39YouTube is the review surface, not the renderer.
PROPOSED989.00 → 1020.00s
LANDED989.00 → 1020.39s
DRIFT+0.00 / +0.39s
RULESword / silence
0of 200
cut points off a word edge
+0ms
measured cut error
274tests
no-network verification suite
100ms
player range enforcement
01 / THE PREMISE
The model is good at judging which moment matters. It is not allowed to guess timing, arithmetic, deduplication, encoding, or whether the finished file is technically sound.
ClipJam scopes probabilistic judgement to the one place it earns its keep—and surrounds it with deterministic gates that can be inspected, repeated, and measured.
02 / THE WORKBENCH
Evidence before verdict.
The review page is not a dashboard painted over model output. It is an audit surface for the exact place a proposed moment became a real cut.
VIDEO JAMMINGCLIPJAM REVIEW
8 CANDIDATES2 APPROVED62.6s APPROVED TIME
ALL CHANGES WRITE TO CLIPS.JSONWhy they called it the Empire of the GreeksLex Fridman Podcast #498 · 16:29 source time
PROPOSEDYOUTUBE PLAYER · REVIEW SURFACE ONLYThe source remains untouched.
Replay the proposed range, then inspect where deterministic snapping landed.
CUT BOUNDARY RULERorange proposal · blue landed range · amber hook window
PROPOSED 989.00 → 1020.00sLANDED 989.00 → 1020.39sDRIFT +0.00 / +0.39sRULES word / silence
APPROVEREJECT↺ REPLAYIN −¼OUT +¼16:9CAPTIONS OFF
SELECTION RATIONALE
Added by hand in manual mode.
REVIEWER NOTES
Keep the breath before the first sentence.
SIGNATURE AUDIT TRAILModel proposes. Deterministic code lands the blade.
03 / THE PIPELINE
Four verbs. Fixed artifacts.
Each stage reads a file and writes the next one. Stop anywhere. Inspect everything. Hand-edit an artifact and resume.
- 01Gemini
Propose
Scan transcripts cheaply, then spend video tokens only on the moments that survive. Or paste timestamps manually with no API key at all.
- 02Deterministic
Land
Snap every proposed edge to a sentence, silence, or word boundary inside a strict drift budget. The model never chooses the final frame.
- 03Human
Review
Approve, reject, nudge, retitle, annotate, and order clips in a local cutting room where every edit writes straight to clips.json.
- 04ffmpeg
Render
Cut real media, normalize geometry, burn captions, stitch the compilation, then verify duration, first-frame luma, and A/V continuity.
sources.jsoncandidates.jsonclips.jsonclips/*.mp4compilation.mp404 / THE DIFFERENCE
Nothing important is hidden.
The file is the state
There is no browser session to drift and no save button to forget. Every mutation is a read-modify-write of clips.json under a lock.
Verification is part of the product
A cut is not done because it looks plausible. ClipJam checks its duration, geometry, first frame, audio/video drift, and frame continuity.
Local-first by construction
Python, FastAPI, JSON, yt-dlp, and ffmpeg. No account, database, cloud renderer, frontend build pipeline, or opaque project file.
05 / THE OUTPUT
A real file, not a screen recording.
ClipJam cuts source media with ffmpeg, normalizes every clip before concatenation, and can produce vertical 9:16 output with captions burned from the same word onsets that informed the boundary.
THE LAST TWO COMMANDS
uv run clipjam cutuv run clipjam render- duration within 100ms
- first frame above black threshold
- audio/video drift within tolerance
- frame timeline continuous at joins
Keep the judgement. Remove the guesswork.
ClipJam is a working local tool and an argument for a particular kind of AI product: agents propose, deterministic gates dispose, and the human can see the seam.